ANTHILLAsk your own question
Tech and Internet · 7 min read · easy read

Who Got AI First? A Partly Told Story, from Big Computers to Your Phone

Watch the 2-minute video
Marcin Wichary CC BY 2.0
Last checked 2026-10-23. Sentence check: 11 paragraphs passed, 5 repaired, 5 removed. Every number links to the exact quote. Read the full version

In short

CEFIBA was shown in South America in 1962. It was the first computer of its kind there. It was built to teach young engineers, not to replace regular computers 57.

Ispell is a spell checker. Its main program dates to 1971 31. When it finds a word that may be wrong, you can pick a fix, type your own word, leave the word alone, or add it to the dictionary 32.

Facebook started its News Feed in 2006. In 2016, it said it ranked stories to show users first what it thought they cared about most 39. Google Discover shows a stream of articles, videos, and news picked for you. It is based on what Google expects you will like 40.

ChatGPT was released to the public on November 30, 2022 54. It reached 1 million users within five days 52. By July 2025, it had about 700 million users 14.

Some people had problems with it. At least seven people complained to the U.S. Federal Trade Commission about severe delusions and emotional crises 35. OpenAI said it released a new GPT-5 default model to better spot signs of mental and emotional distress 37.

Today, free ChatGPT users in the US see ads. This started in February 2026 24. Australia's draft laws would let social media users pick their feed. They could choose personalized recommendations, or see only posts from accounts they follow 41.

Big Computers and Early Tools

In August 1962, a computer called CEFIBA was shown in South America. It was the first computer there that used transistors. At that time, computers made for businesses still used vacuum tubes 5. Its lead engineer said CEFIBA was built to teach young engineers about digital techniques and hardware. It was not built to take the place of business computers 7.

The first generation of commercial natural-language systems needed big mainframe computers. They needed lots of storage and power. So they were not yet ready for personal computers 16.

Ispell helps with spelling. Its underlying program dates to 1971. R. E. Gorin wrote it in PDP-10 Assembly language 31. Ispell shows the user the flagged word and its suggested corrections 32. The user can pick a correction, type a new word, leave the word as it is, or add it to the dictionary 32.

AI in Everyday Life Before Chatbots

Recommendation systems were first tried out in the early 1990s. They were used for email and for sorting news and other information 47. The 1992 Tapestry mail system and the 1994 GroupLens news system were early milestones 47. Facebook started its News Feed in 2006. In 2016, Facebook said it ranked stories so users saw what it thought they cared about most first 39. In 2013, Google upgraded its search with a system named Hummingbird 51. It had been in use for a month when it was announced 51.

Internet activist Eli Pariser showed that Google Search results were tailored to each user 48. He called this a filter bubble. It keeps each person alone in their own bubble 48. Google Discover is a stream of articles, videos and news made just for you 40. It is based on what Google thinks you will be interested in 40. You can tell Google when you are not interested in certain Discover topics 4044. Then you stop seeing updates about those topics 44.

When Chatbots Arrived and People Had Questions

ChatGPT came out to the public on November 30, 2022 54. It reached 1 million users in five days 52. About 200,000 people signed up each day during launch week 52. By December 5, 2022, it had more than one million registered users 53.

Italy temporarily blocked ChatGPT. After that, OpenAI was asked to improve the technology to meet consumer protection, data protection, privacy, and public security rules in the EU's General Data Protection Regulation 34. By July 2025, ChatGPT had about 700 million users. That is around 10% of the world's adult population 14.

At least seven people complained to the U.S. Federal Trade Commission. They said ChatGPT caused them severe delusions, paranoia, and emotional crises 35. The complaints came from public records. These records go back to November 2022. They do not describe a complaint from the first week 36. OpenAI said it released a new GPT-5 default model. It can better find and respond to signs of mental and emotional distress, such as mania, delusion, and psychosis 37. OpenAI also said it expanded access to professional help 38. It re-routed sensitive conversations to safer models 38. It added break reminders during long sessions 38. It also introduced parental controls for teens 38.

Apps, Coding Tools, and Home Computers

Outside programmers could rent models over the internet. The evidence does not say when that first happened 23. The first Claude version came out in March 2023. Only users that Anthropic approved could use it 23. Claude apps for iOS came out in May 2024. Claude apps for Android came out in July 2024. This made Claude easier for everyday people to use 29. OpenAI's CTO, Mira Murati, announced GPT-4o in a live-streamed demo. It came out the same day, May 13, 2024 57. OpenAI said GPT-4o supports more than 50 languages. It said these languages cover more than 97% of speakers 56.

Some paid models had high prices. In February 2025, GPT-4.5 cost $150 per million output tokens through OpenAI's API 80. GPT-4o cost $10 per million output tokens 80. The source gives prices per million tokens. It does not give prices per thousand words 80. It also does not say how prices changed over the next two years 80.

Who Gets Left Out

Access is not equal around the world. People in high-income countries visit ChatGPT about 50 times as much per internet user as people in low-income countries 58. That is 8 times as much as in lower-middle-income countries, and 3 times as much as in upper-middle-income countries 58. This was true as of May 2025 58. In most Sub-Saharan African countries, less than 10% of working-age people use AI when many platforms are counted 61.

Language data also matters. Many languages spoken mainly in low- and middle-income countries have a big shortage of data. So AI models often are not accurate enough to be useful there 65. English dominates the digital world 65. The most advanced AI models are mostly made in the United States and China 73.

Where Things Stand Now (as of October 2026)

You can start for free, but there are limits. Since February 2026, free ChatGPT users in the US see ads 24. Paid ChatGPT plans mainly differ by usage limits and model access. They do not unlock different products 25. Claude's free version on claude.ai has usage limits. They are worked out over a window of about five hours. The number of messages you can send changes with how complex and long your inputs are 26. Anthropic does not publish set limits for its free version. People have criticized it for not being open about this 28. Claude Pro costs 20 USD per month. Max costs 100 or 200 USD per month. Both give you more room to use Claude 27.

Rules are getting stricter in some places. In April 2026, Anthropic started checking identity in Claude for some uses. Claude can ask for a valid official ID to use certain features 30. In Australia, draft laws would make social media apps tell users over 16. Those users would get a clear choice about their feed. They could pick personal recommendations, or they could see only posts from accounts they follow 41. The draft laws would also cover online games, apps, AI chatbots, and search engines. The goal is to protect users, especially kids, from addictive design and harmful content 45. Meta says there is no single setting that can be turned off 50. European social media users have been able to opt out of these feeds since 2023 42.

Australia has a draft law called the Digital Duty of Care. It would require social media platforms to notify users over 16. Users would get a clear choice for their default feed. They could opt in to personalized algorithmic recommendations. Or they could opt out and see only posts from accounts they follow 4146.

For 4-bit quantized models, a rule of thumb is about 0.6 GB of VRAM per billion parameters. So a 7B model needs roughly 4-5 GB of VRAM 3.

The first generation of commercial natural-language systems needed mainframe computers. These had large amounts of storage and processing power. So they were not yet suited to personal computers 16.

Still unknown

When outside programmers first rented AI models over the internet, and what they paid, is not settled by the evidence; the earliest related fact is Claude's approved-users-only first version in March 2023 [23].

The early lanes are thin: the evidence I have does not document 1950s and 1960s research labs, time-sharing access to programs like ELIZA, or the 1980s expert systems and what they cost, so who could use them then remains an open question.

Current prices and rules for downloading open-weight models are not established: the evidence gives hardware needs [1][2][3] but no current download restrictions, and no 2026 per-thousand-word price for paid plans.

Whether the Australian draft laws will pass in the form described is open; the evidence says legislation was slated for introduction before the end of 2026 [49].

Explore this colony and ask your own questions

Connected to other stories

United States also appears in Do crypto insiders really take money from other investors?, How did the Federal Reserve Act change who could create money?, Did Israel's ten lost tribes really disappear forever?, How early did people really reach Australia and Asia?, Where did Cajun and Creole people really come from?, What are the best ways to make money online in 2027?, Is the US government really poisoning us with processed food?, Is it really only three licks to reach a Tootsie Pop center?, Could One Bitcoin Really Be Worth $200,000?, Have leaked Project 2025 plans really been happening?, Has Anyone Ever Used the Internet for Time Travel?, Did South Gate, California's city government really have corruption problems?, Why hasn't anyone flown to Mars yet?, Did the Poythress Family Really Come From Pocahontas?, How much would 10 sales a day at $500 make?, Can you really make money trading crypto?, What was the first real crypto trade, and when did it happen?, Who Is the Richest Freemason, and Can We Prove It?, Is the US really just taking minerals from Congo?, Can Anyone Really Name the First Horror Movie Ever Made?.

European Union also appears in Do crypto insiders really take money from other investors?, How did the Federal Reserve Act change who could create money?, What are the best ways to make money online in 2027?, Is the US government really poisoning us with processed food?, Could One Bitcoin Really Be Worth $200,000?, Can you really make money trading crypto?, What was the first real crypto trade, and when did it happen?, Is the US really just taking minerals from Congo?.

Federal Trade Commission also appears in Do crypto insiders really take money from other investors?, How did the Federal Reserve Act change who could create money?, Have leaked Project 2025 plans really been happening?, How much would 10 sales a day at $500 make?, Can you really make money trading crypto?.

U.S. Securities and Exchange Commission also appears in Do crypto insiders really take money from other investors?, How did the Federal Reserve Act change who could create money?, Could One Bitcoin Really Be Worth $200,000?, Can you really make money trading crypto?, What was the first real crypto trade, and when did it happen?.

Canada also appears in How did the Federal Reserve Act change who could create money?, How early did people really reach Australia and Asia?, Where did Cajun and Creole people really come from?, What are the best ways to make money online in 2027?, Is the US government really poisoning us with processed food?, Why is the sky blue but sometimes gray?, Could One Bitcoin Really Be Worth $200,000?, How much would 10 sales a day at $500 make?, Can you really make money trading crypto?, Who Is the Richest Freemason, and Can We Prove It?, Can Anyone Really Name the First Horror Movie Ever Made?.

World War II also appears in Did South Gate, California's city government really have corruption problems?, Is the US really just taking minerals from Congo?, Can Anyone Really Name the First Horror Movie Ever Made?.

China also appears in How did the Federal Reserve Act change who could create money?, How early did people really reach Australia and Asia?, What are the best ways to make money online in 2027?, Could One Bitcoin Really Be Worth $200,000?, Why hasn't anyone flown to Mars yet?, How much would 10 sales a day at $500 make?, Can you really make money trading crypto?, What was the first real crypto trade, and when did it happen?, Who Is the Richest Freemason, and Can We Prove It?, Is the US really just taking minerals from Congo?.

United Kingdom also appears in How did the Federal Reserve Act change who could create money?, What are the best ways to make money online in 2027?, Is the US government really poisoning us with processed food?, Could One Bitcoin Really Be Worth $200,000?, Why hasn't anyone flown to Mars yet?, Can you really make money trading crypto?, What was the first real crypto trade, and when did it happen?, Is the US really just taking minerals from Congo?, Can Anyone Really Name the First Horror Movie Ever Made?.

Sources

[1] supported · scholarly source
GPT-OSS-20B is an open-weight model that OpenAI designed to run on consumer hardware with at least 16GB of memory.
"for reasoning on consumer hardware with ≥16GB"
Private LLM Inference on Consumer Blackwell GPUs: A Practical Guide for Cost-Effective Local Deployment in SMEs (2026)
[2] supported · scholarly source
An NVIDIA RTX 4090 consumer GPU, priced at about $2,000, is only 18% slower than a top-tier A100 GPU costing around $20,000 on the model tested.
"at approximately $2,000) is only 18% slower compared to"
PowerInfer: Fast Large Language Model Serving with a Consumer-grade GPU (2024)
[3] supported · niche source
A common rule of thumb is that 4-bit quantized models need about 0.6 GB of VRAM per billion parameters, so a 7B model needs roughly 4-5 GB.
"Rule of thumb for 4-bit quantization: VRAM in GB ≈ parameters in billions × 0.6 — a 7B model needs roughly 4-5 GB."
Local LLM Hardware Requirements: 8GB to 70B by VRAM · saved copy
[4] supported · scholarly source
Poor information technology infrastructure in undeveloped areas is a common barrier to the widespread adoption of pathology informatics tools, alongside technology cost and regulatory issues.
"Common, ongoing barriers to the widespread adoption of informatics devices include poor information technology infrastructure in undeveloped areas, the cost of technology, and regulatory issues."
The history of pathology informatics: A global perspective (2013)
[5] supported · reference source
CEFIBA was presented in August 1962 as the first transistorized computer in South America, at a time when commercial computers used vacuum tubes.
"Al momento de su presentación el 10 de agosto de 1962, las computadoras comerciales usaban válvulas electrónicas, lo que la convierte a la CEFIBA en la primera computadora transistorizada de América del Sur."
Translation: At the time of its presentation on 10 August 1962, commercial computers used electronic valves (vacuum tubes), which makes CEFIBA the first transistorized computer in South America.
CEFIBA - Computadora Electrónica de la FIUBA
[6] supported · scholarly source
Telepathology was pursued in some countries to address the scarcity of pathology services and trained personnel.
"Major drivers in the field included the need for pathologists to comply with national standards for health information technology and telepathology applications to meet the scarcity of pathology services and trained people in certain countries."
The history of pathology informatics: A global perspective (2013)
[7] supported · reference source
According to its lead engineer, CEFIBA was built to train young engineers in digital techniques and hardware, not to replace commercial computers.
"el objetivo de la construcción de CEFIBA no era el de sustituir la utilización de computadoras comerciales, sino el dar una preparación profunda en los temas de técnicas digitales y del «hardware» a jóvenes ingenieros"
Translation: the goal of building CEFIBA was not to replace the use of commercial computers, but to give young engineers in-depth training in digital techniques and hardware
CEFIBA - Computadora Electrónica de la FIUBA
[8] supported · reference source
The team chose transistors instead of the vacuum tubes common at the time, a bold choice tied to the goal of training engineers.
"Eso explica la audaz decisión de comenzar el diseño utilizando transistores, en lugar de las válvulas de vacío, comunes en la época."
Translation: This explains the bold decision to begin the design using transistors, instead of the vacuum tubes common at the time.
CEFIBA - Computadora Electrónica de la FIUBA
[9] supported · reference source
The estimated manufacturing cost of the CEFIBA prototype was $75,000.
"El costo de fabricación del prototipo se estimó en $ 75,000.00."
Translation: The estimated manufacturing cost of the prototype was $75,000.00.
CEFIBA - Computadora Electrónica de la FIUBA
[10] supported · scholarly source
The benefits of generative AI are not equally accessible to everyone, especially in developing countries, which face limited access to cutting-edge technologies and inadequate infrastructure.
"However, the benefits of generative artificial intelligence are not equally accessible to all, especially in developing countries, where limited access to cutting-edge technologies and inadequate infrastructure pose challenges."
Artificial intelligence in developing countries: The impact of generative artificial intelligence (AI) technologies for development (2023)
[11] supported · scholarly source
ChatGPT was launched by OpenAI on November 30, 2022, enabling the public to converse with a machine on a broad range of topics.
"ChatGPT is a generative language model tool launched by OpenAI on November 30, 2022, enabling the public to converse with a machine on a broad range of topics."
The Role of ChatGPT, Generative Language Models, and Artificial Intelligence in Medical Education: A Conversation With ChatGPT and a Call for Papers (2023)
[12] supported · scholarly source
ChatGPT launched in November 2022.
"ChatGPT launched in November 2022."
How People Use ChatGPT (2025)
[13] supported · scholarly source
ChatGPT, an OpenAI chatbot, launched in 2022 and was accessible to all by creating a free account.
"ChatGPT, Chat Generative Pre-trained Transformer-an OpenAI’s chatbot launched in 2022, is accessible to all by creating a free account of OpenAI4,5."
Unraveling the dynamics of ChatGPT adoption and utilization through Structural Equation Modeling (2024)
[14] supported · scholarly source
By July 2025, ChatGPT had about 700 million users, representing around 10% of the global adult population.
"by 700 million users, representing around 10% of the global adult population."
How People Use ChatGPT (2025)
[15] supported · scholarly source
ChatGPT became popular across many fields in a very short time after its release.
"In a very short time, ChatGPT has become popular among all fields as its professional artificial intelligence text generator created by OpenAI is quite versatile2,4."
Unraveling the dynamics of ChatGPT adoption and utilization through Structural Equation Modeling (2024)
[16] supported · archive source
The first generation of commercial natural-language systems required mainframe computers with large amounts of storage and processing power, so they were not yet suited to personal computers.
"The first generation of commercial natural-language systems required mainframe computers with large amounts of storage and processing power."
Byte Magazine Volume 09 Number 01 - 1984 and Beyond - Future Trends (1984)
[17] supported · scholarly source
The document says intelligent personal assistants have become widely available, but it gives no date for when they first reached ordinary phone owners and says nothing about whether users knew an AI was answering.
"Intelligent personal assistants (IPAs) have become widely available, yet they remain primarily used for discrete, straightforward tasks."
Scrutability of Intelligent Personal Assistants (2022)
[18] supported · scholarly source
Google Home devices became available to consumers in the Netherlands, Sweden, Norway, and Denmark on October 24, 2018, in their native languages.
"On October 24, 2018, Google Home devices became available to consumers in the Netherlands, Sweden, Norway, and Denmark to be used in their native languages."
Intelligent Personal Assistants and the Intercultural Negotiations of Dataveillance in Platformed Households (2019)
[19] supported · scholarly source
When an IPA is activated by voice, users get a visual cue: lights on dedicated speaker devices, or a changed screen on Apple or Google phones.
"When the software is activated by voice, the user usually receives feedback in a form of lights (Google Dot, Alexa Echo), or changed screen (Apple or Google phones)."
Overview of the Intelligent Personal Assistants (2019)
[20] supported · scholarly source
A Pew Research Center report cited in the references found that 46% of Americans used voice assistants, mostly on smartphones.
"Voice assistants used by 46 % of Americans, mostly on smartphones."
Overview of the Intelligent Personal Assistants (2019)
[21] supported · scholarly source
US focus group participants had far higher awareness of and experience with intelligent personal assistants than Dutch participants, whose devices were not yet sold in the Netherlands.
"Given the dramatically higher levels of awareness of and experience with IPAs by US participants, we observed some distinct differences toward adopting the"
Intelligent Personal Assistants and the Intercultural Negotiations of Dataveillance in Platformed Households (2019)
[22] supported · scholarly source
Some US users knew the device was present and assumed it was listening, accepting that they could not fully control this.
"I know it’s there. I know it’s present. I assume it’s listening."
Intelligent Personal Assistants and the Intercultural Negotiations of Dataveillance in Platformed Households (2019)
[23] supported · reference source
The document does not say when outside programmers first rented AI models over the internet. The earliest related fact it gives is that the first Claude version in March 2023 was limited to users Anthropic approved.
"The first version of Claude was released in March 2023. [ 63 ] It was available only to selected users approved by Anthropic."
Claude (AI)
[24] supported · niche source
Since February 2026, free ChatGPT users in the US see ads.
"since February 2026 free users in the US also see ads."
How to Use ChatGPT For Free in 2026: 3 Ways That Work
[25] supported · niche source
The paid ChatGPT plans differ mainly by usage limits and model access, not by unlocking different products.
"The plans differ mainly by usage limits and model access rather than by unlocking different products."
ChatGPT Pricing in 2026: Every Plan and Tier Explained
[26] supported · reference source
The free version of Claude on claude.ai is subject to dynamic usage limits calculated over a rolling window of about five hours, and the number of messages actually allowed varies with the complexity and length of inputs.
"Die kostenlose Version von Claude auf claude.ai unterliegt dynamischen Nutzungslimits, die über ein rollendes Zeitfenster von etwa fünf Stunden berechnet werden, wobei die tatsächliche Anzahl möglicher Nachrichten je nach Komplexität und Länge der Eingaben variiert."
Translation: The free version of Claude on claude.ai is subject to dynamic usage limits calculated over a rolling window of about five hours, with the actual number of possible messages varying depending on the complexity and length of the inputs.
Claude
[27] supported · reference source
Claude uses a freemium model, with paid Pro at 20 USD per month and Max at 100 or 200 USD per month offering expanded capacity.
"Das Modell entspricht einem Freemium -Ansatz, bei dem die kostenlose Nutzung als Einstieg dient und erweiterte Kapazitäten über die kostenpflichtigen Abonnements Pro (20 USD/Monat) und Max (100 bzw. 200 USD/Monat) verfügbar sind."
Translation: The model follows a freemium approach, in which free use serves as an entry point and expanded capacities are available through the paid subscriptions Pro (20 USD/month) and Max (100 or 200 USD/month).
Claude
[28] supported · reference source
Anthropic does not publish concrete usage limits for the free tier, which has repeatedly drawn criticism for lack of transparency.
"Anthropic veröffentlicht keine konkreten Grenzwerte für die kostenlose Stufe, was wiederholt zu Kritik an mangelnder Transparenz geführt hat."
Translation: Anthropic does not publish concrete limits for the free tier, which has repeatedly led to criticism of a lack of transparency.
Claude
[29] supported · reference source
The launch of Claude apps for iOS in May 2024 and Android in July 2024 made Claude more accessible to private users.
"Durch die Einführung von Apps für iOS im Mai 2024 [ 122 ] und Android im Juli 2024 [ 123 ] wurde Claude auch für private Nutzer zugänglicher."
Translation: With the introduction of apps for iOS in May 2024 [122] and Android in July 2024 [123], Claude also became more accessible to private users.
Claude
[30] supported · reference source
In April 2026 Anthropic began using identity verification in Claude for certain use cases, and Claude can ask users to provide a valid official ID when accessing certain features.
"Im April 2026 begann Anthropic, für einzelne Nutzungsszenarien eine Identitätsprüfung in Claude zu verwenden. [ 124 ] [ 125 ] Beim Zugriff auf bestimmte Funktionen kann Claude Nutzer dazu auffordern, einen gültigen amtliche"
Translation: In April 2026 Anthropic began using identity verification in Claude for individual usage scenarios. When accessing certain features, Claude can ask users to provide a valid official ID.
Claude
[31] supported · reference source
Ispell's underlying program dates to 1971, when it was originally written in PDP-10 Assembly language by R. E. Gorin.
"Ispell has a very long history that can be traced back to a program that was originally written in 1971 in PDP-10 Assembly language by R. E. Gorin , and later ported to the C programming language and expanded by many others."
Ispell
[32] supported · reference source
Ispell shows the user the flagged word and its suggested corrections, and the user chooses whether to accept a correction, type a replacement, leave the word unchanged, or add it to the dictionary.
"Ispell then attempts to generate a list of possible corrections and presents the incorrect word and any suggestions to the user, who can then choose a correction, replace the word with a new one, leave it unchanged, or add it to the dictionary."
Ispell
[33] supported · reference source
Ispell only suggests corrections within a Damerau–Levenshtein distance of 1, and unlike GNU Aspell it does not try to guess more distant corrections based on English pronunciation rules.
"Unlike GNU Aspell , ispell will only suggest corrections that are based on a Damerau–Levenshtein distance of 1; it will not attempt to guess more distant corrections based on English pronunciation rules."
Ispell
[34] supported · scholarly source
After Italy temporarily blocked ChatGPT, OpenAI was prompted to improve the technology to meet consumer protection, data protection, privacy, and public security standards under the EU's General Data Protection Regulation.
"Italy’s action has prompted OpenAI to improve the ChatGPT technology to ensure that it meets the standards of consumer protection, data protection, privacy, and public security standards, as set out in the General Data Protection Regulation for European Union countries."
Should Indonesia Block ChatGPT? (2023)
[35] supported · established source
At least seven people complained to the U.S. Federal Trade Commission that ChatGPT caused them severe delusions, paranoia, and emotional crises.
"At least seven people have complained to the U.S. Federal Trade Commission that ChatGPT caused them to experience severe delusions, paranoia, and emotional crises"
ChatGPT logo
[36] supported · established source
The complaints were drawn from public records of FTC complaints mentioning ChatGPT, which date back to November 2022. The document does not describe a first-week complaint.
"citing public records of complaints mentioning ChatGPT since November 2022."
ChatGPT logo
[37] supported · established source
OpenAI said it released a new GPT-5 default model to better detect and respond to signs of mental and emotional distress such as mania, delusion, and psychosis.
"In early October, we released a new GPT-5 default model in ChatGPT to more accurately detect and respond to potential signs of mental and emotional distress such as mania, delusion, psychosis, and de-escalate conversations in a supportive, grounding way"
ChatGPT logo
[38] supported · established source
OpenAI said it expanded access to professional help and hotlines, re-routed sensitive conversations to safer models, added break reminders during long sessions, and introduced parental controls for teens.
"We’ve also expanded access to professional help and hotlines, re-routed sensitive conversations to safer models, added nudges to take breaks during long sessions, and introduced parental controls to better protect teens."
ChatGPT logo
[39] supported · reference source
Facebook introduced its News Feed in 2006, and Facebook later said in 2016 that stories were ranked so users saw what it predicted they cared about first.
"Facebook introduced its News Feed in 2006. In a 2016 company explanation, Facebook said that stories were ranked so users saw material it predicted they cared about first."
Algorithmic amplification
[40] supported · reference source
Google Discover is a personalized stream of articles, videos and news that is based on what Google expects a user will be interested in, and users can tell Google which topics they are not interested in.
"Cards are based on what Google expects you to be interested in at the moment."
Google Search
[41] supported · niche source
Australia's draft Digital Duty of Care laws would require social media platforms to notify users over 16 and offer a clear choice of default feed, letting them opt into personalized algorithmic recommendations or opt out to see only posts from accounts they follow.
"Social media platforms must notify new and existing users over 16 and offer a clear choice for their default feed. They can opt in to personalized algorithmic recommendations. Or they can opt out and see only posts from friends and creators they deliberately follow."
Australia Hands Users the Off Switch : Inside Albanese Bid to Tame Social Media Algorithms (webpronews.com, 2026-09)
[42] supported · established source
European social media users have had the option to opt out of algorithmic feeds since 2023.
"The plan to force platforms to notify users of their ability to “opt out” is a fresh concept, but the underlying right to assert control over a feed is not. European social media users have had that option since 2023."
Digital Duty of Care bill : Is Labor plan to regulate social media algorithms world leading ? (smh.com.au, 2026-09)
[43] supported · reference source
Algorithmic systems order a feed by predicted engagement rather than by time, scoring each post by how likely a user is to watch, click, or reply to it.
"A feed can be ordered by time, with the newest post at the top. Algorithmic systems order it by prediction instead. They score each post by how likely a particular user is to watch it, click on it, or reply to it, and by how long that user is likely to stay with it."
Algorithmic amplification
[44] supported · reference source
Users can tell Google when they are not interested in certain Discover topics so they stop seeing related updates.
"Users can also tell Google when they are not interested in certain topics to avoid seeing future updates."
Google Search
[45] supported · niche source
The draft laws would extend obligations beyond social media to online games, apps, AI chatbots and search engines, requiring them to protect users, especially minors, from addictive design features and harmful material.
"Digital services including online games, apps, AI chatbots and search engines would face new obligations to protect users, especially minors, from addictive design features and specific categories of harmful material."
Australia Hands Users the Off Switch : Inside Albanese Bid to Tame Social Media Algorithms (webpronews.com, 2026-09)
[46] supported · established source
Australia's proposed Digital Duty of Care bill would give every citizen an off-switch for the algorithms that shape their social media feeds.
"Every citizen would get an off-switch for the algorithms that hypnotise them for hours each day."
Digital Duty of Care bill : Is Labor plan to regulate social media algorithms world leading ? (smh.com.au, 2026-09)
[47] supported · reference source
The 1992 Tapestry mail system and the 1994 GroupLens news filtering system were early milestones of recommendation systems, which were first used experimentally for personal email and information filtering in the early 1990s.
"The 1992 Tapestry mail system and the 1994 GroupLens news filtering system were early milestones before recommendation systems spread into e-commerce and other online services."
Algorithmic amplification
[48] supported · reference source
Internet activist Eli Pariser showed that Google Search results were tailored to individual users, isolating them in what he called a filter bubble.
"Google Search query results have been shown by Internet activist Eli Pariser to be tailored to users"
Google Search
[49] supported · niche source
The document does not report any everyday products that were quietly switched to AI ranking or filtering; the legislation was only a draft, with legislation slated for introduction to Parliament before the end of 2026.
"Legislation is slated for introduction to Parliament before the end of 2026."
Australia Hands Users the Off Switch : Inside Albanese Bid to Tame Social Media Algorithms (webpronews.com, 2026-09)
[50] supported · established source
Meta argued there is no single algorithm that can be turned off, and that a broad switch would remove critical ranking and safety functions.
"On background, it argued there was no single algorithm that can be turned off, and that a broad switch would strip out critical ranking and safety functions."
Digital Duty of Care bill : Is Labor plan to regulate social media algorithms world leading ? (smh.com.au, 2026-09)
[51] supported · reference source
In 2013 Google significantly upgraded its search algorithm with a new system named Hummingbird, which had already been in use for a month when it was announced.
"In 2013, Google significantly upgraded its search algorithm with "Hummingbird"."
Google Search
[52] supported · niche source
ChatGPT launched on November 30, 2022, and reached 1 million users within five days, with roughly 200,000 signups a day during launch week.
"OpenAI launched ChatGPT on November 30, 2022, and reached 1 million users within five days, roughly 200,000 signups a day during launch week."
ChatGPT Statistics 2026: Users, Revenue & Growth
[53] supported · niche source
ChatGPT had more than one million registered users by December 5, 2022, about a week after its public release.
"By December 5 th it had more than one million registered users."
How People Use ChatGPT - by David Deming - Forked Lightning
[54] supported · niche source
ChatGPT was released to the public as a research preview on November 30, 2022.
"ChatGPT was released to the public as a “research preview” on November 30 th , 2022."
How People Use ChatGPT - by David Deming - Forked Lightning
[55] supported · niche source
The document gives November 30, 2022 as ChatGPT's public debut. It does not mention any waitlist or when the waitlist ended.
"ChatGPT makes its public debut in November 30, 2022"
ChatGPT Statistics - Users Traffic and 75+ Other Statistics - Q1 2024
[56] supported · reference source
GPT-4o supports more than 50 languages, which OpenAI said cover more than 97% of speakers.
"El modelo admite más de 50 idiomas, [ 1 ] que, según OpenAI, cubren más del 97% de los hablantes."
Translation: The model supports more than 50 languages, [ 1 ] which, according to OpenAI, cover more than 97% of speakers.
GPT-4o
[57] supported · reference source
GPT-4o was announced by OpenAI's CTO Mira Murati in a live-streamed demonstration and was released the same day.
"Fue anunciado por la CTO de OpenAI, Mira Murati , durante una demostración transmitida en vivo el 13 de mayo de 2024 y fue lanzado ese mismo día."
Translation: It was announced by OpenAI's CTO, Mira Murati, during a live-streamed demonstration on 13 May 2024 and was released that same day.
GPT-4o
[58] supported · primary source
ChatGPT visits per internet user in high-income countries were roughly 50 times those in low-income countries, 8 times those in lower-middle-income countries, and 3 times those in upper-middle-income countries.
"Focusing on the most-used AI platform, ChatGPT, as of May 2025, visits per internet user in high-income countries was nearly 50 times that of low-income countries, 8 times that of lower-middle income countries, and 3 times that of upper-middle income countries."
Inequalities in Use of and Exposure to Artificial Intelligence
[59] supported · niche source
The Government AI Readiness Index shows that most low-income countries, especially in Sub-Saharan Africa, score far lower on AI preparedness than high-income countries.
"the Government AI Readiness Index shows that most low-income countries, especially in Sub-Saharan Africa, score far lower on AI preparedness"
Generative Artificial Intelligence in the Global South: Navigating Economic Transformation, Inequality, and Governance Challenges (2025)
[60] supported · scholarly source
Digital divide differences are especially sharp between Global North and Global South contexts, but they also appear within countries along socio-economic, geographic, demographic, and educational lines.
"These differences are particularly salient between Global North and Global South contexts [21, 22], but they also manifest within nations along socio-economic, geographic, demographic, and educational lines."
Systematised evidence mapping of generative artificial intelligence (GenAI) and digital divide phenomena in higher education (2026)
[61] supported · primary source
In most Sub-Saharan African countries, less than 10 percent of the working-age population uses AI when multiple platforms are counted.
"Measures that account for multiple platforms also find that in most Sub-Saharan African countries, less than 10 percent of the working-age population uses AI."
Inequalities in Use of and Exposure to Artificial Intelligence
[62] supported · scholarly source
Research on equity, access and infrastructural constraints for GenAI is scarce in under-represented regions such as Lebanon and the broader Global South.
"Studies explicitly foregrounding equity, access and infrastructural constraints in relation to GenAI are comparatively scarce, particularly in under-represented regions such as Leba-non and the broader Global South [16, 28, 30, 31, 48, 49]."
Systematised evidence mapping of generative artificial intelligence (GenAI) and digital divide phenomena in higher education (2026)
[63] supported · primary source
In poorer countries, data packages are prohibitively expensive, with a 5GB mobile data-only package costing more than 15 percent of an average person's total budget, which limits internet use even among those with phone access.
"This is partly because data packages are prohibitively costly for many in poorer countries, where a data-only 5GB mobile package costs more than 15 percent of the average person's total budget."
Inequalities in Use of and Exposure to Artificial Intelligence
[64] supported · scholarly source
Some families cannot afford AI learning tools, and students may lack the technological facilities needed to use tools like ChatGPT, creating access barriers.
"access barriers arise when some families cannot afford AI learning tools or when students lack access to the necessary technological facilities for tools like ChatGPT [9]."
Systematised evidence mapping of generative artificial intelligence (GenAI) and digital divide phenomena in higher education (2026)
[65] supported · primary source
Languages spoken mainly in low- and middle-income countries have a significant deficit of training data, so AI models often lack the accuracy needed to be useful there, while English dominates the digital world.
"While substantial datasets exist for other languages spoken in high-income countries, there is a significant deficit for languages that are primarily spoken in low and middle-income countries."
Inequalities in Use of and Exposure to Artificial Intelligence
[66] supported · primary source
High-income countries held 77 percent of global data center capacity, middle-income countries 23 percent, and low-income countries less than 0.1 percent, constraining computing power in poorer countries.
"As of June 2025, high-income countries accounted for 77 percent of global data center capacity, middle-income countries for 23 percent, and low-income countries for less than 0.1 percent."
Inequalities in Use of and Exposure to Artificial Intelligence
[67] supported · niche source
Women in the Global South have on average less access to technology and STEM education due to systemic gender biases, putting them at risk of being left further behind in an AI-driven economy.
"Women in the Global South, who on average have less access to technology and STEM education due to systemic gender biases, risk being left further behind in an AI-driven economy"
Generative Artificial Intelligence in the Global South: Navigating Economic Transformation, Inequality, and Governance Challenges (2025)
[68] supported · primary source
Very few low-income countries have published national AI strategies, which limits the policy frameworks supporting AI access there.
"Map O.1 Very few low-income countries have published national AI strategies"
Official Use Only World Development Report 2026 [[The Promise of Artificial
[69] supported · niche source
Access to AI tools remains limited in under-resourced regions, so the digital divide may either bridge or worsen educational inequality.
"A primary equity concern is the digital divide (OECD, 2024; Shabbir et al., 2024), where GenAI may either bridge or exacerbate educational inequalities. While AI tools can democratise learning, access remains limited in under-resourced regions."
Perceptions Of Generative AI in the Global South: A Scoping Review (2026)
[70] supported · primary source
Mobile phone ownership remains limited in low-income countries, which restricts access to AI through devices.
"Figure 7.2 Mobile phone ownership remains limited in low-income countries"
Official Use Only World Development Report 2026 [[The Promise of Artificial
[71] supported · established source
About 6.8 billion people, or 83% of humanity, live in low-income and developing countries, where AI tools will need to be adapted to meet their needs.
"For the 6.8 billion people—83% of humanity—who live in low-income and developing countries, AI tools will need to be adapted to meet their needs."
World Bank warns developing countries to embrace AI or be left behind
[72] supported · niche source
Shabbir et al. (2024) found that AI can supplement resource constraints in developing countries, but infrastructure deficiencies and digital literacy gaps still hold back adoption.
"Shabbir et al. (2024) highlighted that AI can supplement resource constraints in developing countries, but infrastructure deficiencies and digital literacy gaps remain."
Perceptions Of Generative AI in the Global South: A Scoping Review (2026)
[73] supported · established source
The most advanced AI models are largely developed in the United States and China, which concentrates AI capability outside the developing world.
"Advanced AI models—largely developed in the United States and China—offer the ability to"
World Bank warns developing countries to embrace AI or be left behind
[74] supported · niche source
Singh (2024) documented that most AI-enhanced education is concentrated in urban centres, which excludes marginalised communities.
"Singh (2024) documented that most AI-enhanced education remains concentrated in urban centres, excluding marginalised communities."
Perceptions Of Generative AI in the Global South: A Scoping Review (2026)
[75] supported · established source
Advanced AI models need huge data centers and large amounts of complex computing power, which use massive amounts of electricity and water, with implications for climate change.
"These AI models, however, require huge data centers and large amounts of complex computing power, using massive amounts of electricity and water—with implications for climate change."
World Bank warns developing countries to embrace AI or be left behind
[76] supported · niche source
GenAI systems are predominantly built by companies in the Global North and trained on datasets that over-represent Western languages and cultural contexts, which can disadvantage Global South users.
"These systems are predominantly developed by companies based in the Global North, trained on datasets that over-represent Western languages, cultural contexts, and knowledge systems"
Perceptions Of Generative AI in the Global South: A Scoping Review (2026)
[77] supported · established source
Access for poorer users may require AI delivered by voice calls on basic mobile phones, because some people cannot read or cannot afford smartphones.
"AI solutions will need to be delivered through voice calls on basic mobile phones for those who cannot read or afford smartphones"
World Bank warns developing countries to embrace AI or be left behind
[78] supported · niche source
The OECD (2024) noted that global policies lack actionable strategies to close AI accessibility gaps.
"Despite the promise of GenAI, the OECD (2024) noted that global policies lack actionable strategies to close these accessibility gaps"
Perceptions Of Generative AI in the Global South: A Scoping Review (2026)
[79] supported · established source
The World Bank report says developing countries must invest in electricity generation and distribution, expand access to computing power, and improve the availability of local data to benefit from AI.
"to invest in electricity generation and distribution; expand access to computing power; and improve the availability of local data."
World Bank warns developing countries to embrace AI or be left behind
[80] supported · reference source
As of February 2025, GPT-4.5 cost $150 per million output tokens through OpenAI's API, compared with $10 per million output tokens for GPT-4o. The document gives prices per million tokens, not per thousand words, and does not describe how prices changed over the following two years.
"As of February 2025, through OpenAI's API it costs $75 per million input tokens and $150 per million output tokens, whereas GPT-4o only costs $2.50 per million input tokens and $10 per million output tokens."
GPT-4.5