ClaimsSteven Pinker

“these systems have anterograde amnesia for everything that happens after 2021.”

Steven Pinker, interviewee 2023 The Intelligence and Rationality of AI and Humans: A Conversation With Steven Pinker (Harvard Data Science Review) archived copy

Factual AI capabilities

Current large language models are trained on a fixed dataset frozen as of 2021 and have no knowledge of events after 2021.

Mostly true True in substance, wrong in a detail.

Factual: Statements about the present or the past, checked against the record: did they hold up?

What would count as true or false

Judge as of 2023, when the claim was made. The claim is that the prominent LLM systems of the time (e.g., ChatGPT/GPT-3.5, GPT-4) were trained on data ending around 2021 and knew nothing of later events. True: the leading widely used LLMs of 2023 had documented training-data cutoffs in 2021 (e.g., OpenAI's stated September 2021 cutoff) and could not report later events without outside retrieval. Mostly true: the main systems had 2021 cutoffs, but the exact month varied, or a few systems had slightly later cutoffs or limited browsing/retrieval add-ons that the claim leaves out. Mixed: in 2023 roughly as many prominent systems had post-2021 knowledge (later cutoffs or built-in web retrieval, e.g., Bing Chat, Bard) as had 2021 cutoffs. Mostly false: the 2021 cutoff applied only to one product or an early version, while most current LLMs at the time had later cutoffs or live retrieval. False: the main LLMs of 2023 did not have 2021 training cutoffs.

Set before the research began.

Evidence (4)

  1. E1 Can ChatGPT fact-check? PolitiFact tested. PolitiFact · 30 May 2023

    “It does not have access to any data after September 2021, meaning it is blissfully unaware of big global events such as Queen Elizabeth II’s death and Russia’s invasion of Ukraine.”

  2. E2 Anthropic releases Claude 2, its second-gen AI chatbot | TechCrunch TechCrunch · 11 July 2023

    “Claude 2 was trained on more recent data — a mix of websites, licensed data sets from third parties and voluntarily-supplied user data from early 2023, roughly 10% of which is non-English”

  3. E3 Can Base ChatGPT be Used for Forecasting without Additional Optimization? arXiv (ar5iv rendering) · 19 April 2023

    “we take advantage of the fact that the training data at the time of our experiments (April and May 2023) stopped at September 2021”

  4. E4 ChatGPT and Fake Citations - Duke University Libraries Blogs Duke University Libraries · 9 March 2023

    “ChatGPT’s “knowledge” is based on the dataset that was available before September 2021, and therefore, it may not be able to provide up-to-date information on current events or predict the future.”

Reasoning

Several independent 2023 sources say ChatGPT had no training data after September 2021 and did not know about later events (E1, E3, E4). However, E2 reports that Claude 2, released in July 2023, was trained on more recent data that included material from early 2023. That shows the 2021 cutoff did not apply to every prominent system. The core claim fits the main system, but it leaves out models with later cutoffs, which matches the "mostly true" criterion.

Cite this claim

Steven Pinker, “Current large language models are trained on a fixed dataset frozen as of 2021 and have no knowledge of events after 2021.” (said 2023). Verdict: mostly true. Reliability API, claim c_bfbdbf2a930c, method 0.1.0+28da144b, dataset release v1.1.2. https://epistacklabs.com/claims/c_bfbdbf2a930c

This claim is in dataset release v1.1.2, with the same verdict.

Reply to this claim

For Steven Pinker, or someone speaking for them: a correction or counter-evidence. Replies are read before anything is shown; an approved reply shows your name, role, date, statement and link. Your email address is never shown and is deleted 30 days after the decision (privacy).

More factual claims by Steven Pinker

  1. Mostly true Factual 11 October 2026
    “A major fact about language, known for decades, is that people don’t blurt out what they mean”

    It has been known for decades in the study of language that people do not blurt out what they mean but use euphemism, innuendo and indirection.

  2. Mostly true Factual 7 October 2026
    “Computers have always been better at some of the things that humans do, otherwise we never would have adopted them”

    Computers have always been better than humans at some tasks, which is why they were adopted from the 1950s for databases and calculation.

  3. Resolved true Factual 7 October 2026
    “then it is very much unlike AlphaGo, with this closed universe governed by fixed rules.”

    AlphaGo operated in a closed universe governed by fixed rules, unlike the open-loop real world.

  4. Mostly true Factual 26 September 2026
    “The capabilities of LLMs trained on gargantuan datasets are more general than I (and almost everyone else) would have guessed at the time.”

    The capabilities of LLMs trained on huge datasets are more general than Pinker and almost everyone else would have guessed at the time he laid out his arguments, five years before LLM chatbots were released.

  5. Mostly true Factual 24 September 2025
    “There are some features in large language models that actually implicitly build in some of the insight of linguistics.”

    Large language models have features, such as hidden layers, that implicitly build in some insights of linguistics.

All of them

Verdicts are made automatically; how verdicts are made.