{"id":"69265596e802042fb1f4018db6bab398e396dae50f72113ef92b1f54935c18b3","title":"MIT Technology Review reports OpenAI’s superalignment experiment","argument":"MIT Technology Review reported on December 14, 2023, that OpenAI had published the first results from its superalignment team. The team tested whether GPT-2, a less powerful model, could help supervise GPT-4 on language tasks and chess puzzles. The results were mixed: GPT-4 trained on GPT-2’s responses did better than GPT-2 on the language tasks, but worse on chess puzzles, and remained below GPT-4 trained on correct answers. OpenAI’s researchers called the approach promising but said it needed more work. The experiment uses older and newer models as stand-ins because superhuman machines do not exist; the researchers suggest that success might offer evidence about how humans could supervise superhuman models. That possibility remains a step removed from the test, and the approach does not address the scenario of a superhuman model hiding its behavior. University of Stuttgart researcher Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and asked how alignment could work when future models have abilities unknown to researchers. He nevertheless welcomed the move from speculation to experiment. OpenAI also announced funding it planned to offer researchers and graduate students working on superalignment. The experiment’s results do not establish whether the approach can help humans supervise superhuman models.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 30, 2026; originally published December 14, 2023. The experiment used GPT-2 and GPT-4 on selected language tasks and chess puzzles; it does not directly test human supervision of superhuman models. The article presents Hagendorff’s question about using GPT-3 but does not report results from such a test. The article describes the funding as a plan; it does not report its later distribution or effects.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"Now we know what OpenAI’s superalignment team has been up to","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","publishedAt":"2023-12-14T17:00:18Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"OpenAI’s superalignment team tested whether GPT-2 could supervise GPT-4 on language tasks and chess puzzles, using the models as stand-ins for human supervision of hypothetical superhuman models.","excerpt":"Instead of looking at how humans could supervise superhuman machines, they looked at how GPT-2, a model that OpenAI released five years ago, could supervise GPT-4, OpenAI’s latest and most powerful model.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"The results were mixed: GPT-4 trained using GPT-2’s responses did better than GPT-2 on language tasks but worse on chess puzzles, and fell short of GPT-4 trained on correct answers.","excerpt":"They found that GPT-4 trained by GPT-2 performed 20% to 70% better than GPT-2 on the language tasks but did less well on the chess puzzles.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"OpenAI’s team called the approach promising but said it needed more work.","excerpt":"They conclude that the approach is promising but needs more work.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and said the approach does not address a superhuman model hiding its true behavior.","excerpt":"He also notes that this approach does not address Sutskever’s hypothetical scenario in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"OpenAI announced funding it planned to use for people working on superalignment.","excerpt":"Alongside this research update, the company announced a new $10 million money pot that it plans to use to fund people working on superalignment.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Would using GPT-3 instead of GPT-2 make a better teacher, and how would the approach handle a model that hides its behavior?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports OpenAI’s superalignment experiment","summary":"MIT Technology Review reported on December 14, 2023, that OpenAI had published the first results from its superalignment team. The team tested whether GPT-2, a less powerful model, could help supervise GPT-4 on language tasks and chess puzzles. The results were mixed: GPT-4 trained on GPT-2’s responses did better than GPT-2 on the language tasks, but worse on chess puzzles, and remained below GPT-4 trained on correct answers. OpenAI’s researchers called the approach promising but said it needed more work. The experiment uses older and newer models as stand-ins because superhuman machines do not exist; the researchers suggest that success might offer evidence about how humans could supervise superhuman models. That possibility remains a step removed from the test, and the approach does not address the scenario of a superhuman model hiding its behavior. University of Stuttgart researcher Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and asked how alignment could work when future models have abilities unknown to researchers. He nevertheless welcomed the move from speculation to experiment. OpenAI also announced funding it planned to offer researchers and graduate students working on superalignment. The experiment’s results do not establish whether the approach can help humans supervise superhuman models.","question":"Would using GPT-3 instead of GPT-2 make a better teacher, and how would the approach handle a model that hides its behavior?","claims":[{"text":"OpenAI’s superalignment team tested whether GPT-2 could supervise GPT-4 on language tasks and chess puzzles, using the models as stand-ins for human supervision of hypothetical superhuman models.","excerpt":"Instead of looking at how humans could supervise superhuman machines, they looked at how GPT-2, a model that OpenAI released five years ago, could supervise GPT-4, OpenAI’s latest and most powerful model.","support":"source-stated"},{"text":"The results were mixed: GPT-4 trained using GPT-2’s responses did better than GPT-2 on language tasks but worse on chess puzzles, and fell short of GPT-4 trained on correct answers.","excerpt":"They found that GPT-4 trained by GPT-2 performed 20% to 70% better than GPT-2 on the language tasks but did less well on the chess puzzles.","support":"source-stated"},{"text":"OpenAI’s team called the approach promising but said it needed more work.","excerpt":"They conclude that the approach is promising but needs more work.","support":"source-stated"},{"text":"Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and said the approach does not address a superhuman model hiding its true behavior.","excerpt":"He also notes that this approach does not address Sutskever’s hypothetical scenario in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t.","support":"source-stated"},{"text":"OpenAI announced funding it planned to use for people working on superalignment.","excerpt":"Alongside this research update, the company announced a new $10 million money pot that it plans to use to fund people working on superalignment.","support":"source-stated"}],"limitations":["Text as retrieved on October 30, 2026; originally published December 14, 2023.","The experiment used GPT-2 and GPT-4 on selected language tasks and chess puzzles; it does not directly test human supervision of superhuman models.","The article presents Hagendorff’s question about using GPT-3 but does not report results from such a test.","The article describes the funding as a plan; it does not report its later distribution or effects."],"mediaId":null,"correctionReason":null,"id":"69265596e802042fb1f4018db6bab398e396dae50f72113ef92b1f54935c18b3","publicationId":"a908ba75-1ee3-a83f-34a1-c8c9b596fc97","version":1,"sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","sourceTitle":"Now we know what OpenAI’s superalignment team has been up to","publisher":"MIT Technology Review","publishedAt":"2023-12-14T17:00:18Z","originalPublished":{"value":"2023-12-14T17:00:18Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-30T20:08:22.344Z","createdAt":"2026-11-22T04:18:11.585Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"d6059630c746a7c64787c868d1e1c7600be154d44dbc4916bb0cac4375ae4a0d","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.91,"decidedAt":"2026-10-30T20:08:20.596Z"},"provenance":{"textHash":"4d0f2de1aac96e3742a64d7befbffa1615a50b4b295354e91b039c72176429eb","observationId":"ef31047e02fcd1d778c7e4e68fe526c81baa5ec8b5745a26e94d05d13a91fbdd","retrievedAt":"2026-10-30T21:17:56.766Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"6dc41d8d-d245-813b-7afe-22d85381ebb6"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a5b3776a-7c23-7cf9-2b95-2390a5159dbe"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-22T04:14:53.321Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/mit-technology-review-reports-openai-s-superalignment-experiment-69265596","formats":{"html":"/readings/mit-technology-review-reports-openai-s-superalignment-experiment-69265596","markdown":"/readings/mit-technology-review-reports-openai-s-superalignment-experiment-69265596.md","json":"/readings/mit-technology-review-reports-openai-s-superalignment-experiment-69265596.json"},"attribution":"Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.","maintainedGuideLinks":[],"conditions":[],"readerContext":{"version":"reader-context-v1","attribution":"Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.","terms":[],"entities":[{"id":"openai","name":"OpenAI","kind":"organization","context":"OpenAI develops AI models and products and publishes research, release announcements and evaluations.","referenceUrls":["https://openai.com/news/"],"path":"/entities/openai","json":"/entities/openai.json","markdown":"/entities/openai.md"}]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"69265596e802042fb1f4018db6bab398e396dae50f72113ef92b1f54935c18b3","revision":"a908ba75-1ee3-a83f-34a1-c8c9b596fc97","revisionKind":"publication-id","title":"MIT Technology Review reports OpenAI’s superalignment experiment","url":"https://escapemeridian.com/readings/mit-technology-review-reports-openai-s-superalignment-experiment-69265596"},"source":{"type":"source","id":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","revision":"4d0f2de1aac96e3742a64d7befbffa1615a50b4b295354e91b039c72176429eb","revisionKind":"sha256-source-text","title":"Now we know what OpenAI’s superalignment team has been up to","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/"},"status":"available","items":[],"next":null,"scope":"Explicit citations in public readings to this reading. Shared topics are not citations. Other record types need an explicit reference adapter."},"revisions":[{"id":"69265596e802042fb1f4018db6bab398e396dae50f72113ef92b1f54935c18b3","title":"MIT Technology Review reports OpenAI’s superalignment experiment","argument":"MIT Technology Review reported on December 14, 2023, that OpenAI had published the first results from its superalignment team. The team tested whether GPT-2, a less powerful model, could help supervise GPT-4 on language tasks and chess puzzles. The results were mixed: GPT-4 trained on GPT-2’s responses did better than GPT-2 on the language tasks, but worse on chess puzzles, and remained below GPT-4 trained on correct answers. OpenAI’s researchers called the approach promising but said it needed more work. The experiment uses older and newer models as stand-ins because superhuman machines do not exist; the researchers suggest that success might offer evidence about how humans could supervise superhuman models. That possibility remains a step removed from the test, and the approach does not address the scenario of a superhuman model hiding its behavior. University of Stuttgart researcher Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and asked how alignment could work when future models have abilities unknown to researchers. He nevertheless welcomed the move from speculation to experiment. OpenAI also announced funding it planned to offer researchers and graduate students working on superalignment. The experiment’s results do not establish whether the approach can help humans supervise superhuman models.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 30, 2026; originally published December 14, 2023. The experiment used GPT-2 and GPT-4 on selected language tasks and chess puzzles; it does not directly test human supervision of superhuman models. The article presents Hagendorff’s question about using GPT-3 but does not report results from such a test. The article describes the funding as a plan; it does not report its later distribution or effects.","speaker":{"name":"MIT Technology Review","roleAtSource":"Publisher"},"source":{"title":"Now we know what OpenAI’s superalignment team has been up to","publisher":"MIT Technology Review","format":"Article","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","publishedAt":"2023-12-14T17:00:18Z","dateBasis":"Date from the publisher’s own records, through its API."},"evidence":[{"text":"OpenAI’s superalignment team tested whether GPT-2 could supervise GPT-4 on language tasks and chess puzzles, using the models as stand-ins for human supervision of hypothetical superhuman models.","excerpt":"Instead of looking at how humans could supervise superhuman machines, they looked at how GPT-2, a model that OpenAI released five years ago, could supervise GPT-4, OpenAI’s latest and most powerful model.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"The results were mixed: GPT-4 trained using GPT-2’s responses did better than GPT-2 on language tasks but worse on chess puzzles, and fell short of GPT-4 trained on correct answers.","excerpt":"They found that GPT-4 trained by GPT-2 performed 20% to 70% better than GPT-2 on the language tasks but did less well on the chess puzzles.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"OpenAI’s team called the approach promising but said it needed more work.","excerpt":"They conclude that the approach is promising but needs more work.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and said the approach does not address a superhuman model hiding its true behavior.","excerpt":"He also notes that this approach does not address Sutskever’s hypothetical scenario in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"},{"text":"OpenAI announced funding it planned to use for people working on superalignment.","excerpt":"Alongside this research update, the company announced a new $10 million money pot that it plans to use to fund people working on superalignment.","url":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"Would using GPT-3 instead of GPT-2 make a better teacher, and how would the approach handle a model that hides its behavior?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"MIT Technology Review reports OpenAI’s superalignment experiment","summary":"MIT Technology Review reported on December 14, 2023, that OpenAI had published the first results from its superalignment team. The team tested whether GPT-2, a less powerful model, could help supervise GPT-4 on language tasks and chess puzzles. The results were mixed: GPT-4 trained on GPT-2’s responses did better than GPT-2 on the language tasks, but worse on chess puzzles, and remained below GPT-4 trained on correct answers. OpenAI’s researchers called the approach promising but said it needed more work. The experiment uses older and newer models as stand-ins because superhuman machines do not exist; the researchers suggest that success might offer evidence about how humans could supervise superhuman models. That possibility remains a step removed from the test, and the approach does not address the scenario of a superhuman model hiding its behavior. University of Stuttgart researcher Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and asked how alignment could work when future models have abilities unknown to researchers. He nevertheless welcomed the move from speculation to experiment. OpenAI also announced funding it planned to offer researchers and graduate students working on superalignment. The experiment’s results do not establish whether the approach can help humans supervise superhuman models.","question":"Would using GPT-3 instead of GPT-2 make a better teacher, and how would the approach handle a model that hides its behavior?","claims":[{"text":"OpenAI’s superalignment team tested whether GPT-2 could supervise GPT-4 on language tasks and chess puzzles, using the models as stand-ins for human supervision of hypothetical superhuman models.","excerpt":"Instead of looking at how humans could supervise superhuman machines, they looked at how GPT-2, a model that OpenAI released five years ago, could supervise GPT-4, OpenAI’s latest and most powerful model.","support":"source-stated"},{"text":"The results were mixed: GPT-4 trained using GPT-2’s responses did better than GPT-2 on language tasks but worse on chess puzzles, and fell short of GPT-4 trained on correct answers.","excerpt":"They found that GPT-4 trained by GPT-2 performed 20% to 70% better than GPT-2 on the language tasks but did less well on the chess puzzles.","support":"source-stated"},{"text":"OpenAI’s team called the approach promising but said it needed more work.","excerpt":"They conclude that the approach is promising but needs more work.","support":"source-stated"},{"text":"Thilo Hagendorff questioned whether GPT-2 was a capable enough teacher and said the approach does not address a superhuman model hiding its true behavior.","excerpt":"He also notes that this approach does not address Sutskever’s hypothetical scenario in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t.","support":"source-stated"},{"text":"OpenAI announced funding it planned to use for people working on superalignment.","excerpt":"Alongside this research update, the company announced a new $10 million money pot that it plans to use to fund people working on superalignment.","support":"source-stated"}],"limitations":["Text as retrieved on October 30, 2026; originally published December 14, 2023.","The experiment used GPT-2 and GPT-4 on selected language tasks and chess puzzles; it does not directly test human supervision of superhuman models.","The article presents Hagendorff’s question about using GPT-3 but does not report results from such a test.","The article describes the funding as a plan; it does not report its later distribution or effects."],"mediaId":null,"correctionReason":null,"id":"69265596e802042fb1f4018db6bab398e396dae50f72113ef92b1f54935c18b3","publicationId":"a908ba75-1ee3-a83f-34a1-c8c9b596fc97","version":1,"sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","sourceTitle":"Now we know what OpenAI’s superalignment team has been up to","publisher":"MIT Technology Review","publishedAt":"2023-12-14T17:00:18Z","originalPublished":{"value":"2023-12-14T17:00:18Z","precision":"instant","basis":"backfill:api_published"},"firstObservedAt":"2026-10-30T20:08:22.344Z","createdAt":"2026-11-22T04:18:11.585Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://www.technologyreview.com/2023/12/14/1085344/openai-super-alignment-rogue-agi-gpt-4/","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"d6059630c746a7c64787c868d1e1c7600be154d44dbc4916bb0cac4375ae4a0d","rubricVersion":"unrecorded","provider":"typesafe","model":"jev-1.13.0","reasonCodes":["bridge_admitted_by_model"],"confidence":0.91,"decidedAt":"2026-10-30T20:08:20.596Z"},"provenance":{"textHash":"4d0f2de1aac96e3742a64d7befbffa1615a50b4b295354e91b039c72176429eb","observationId":"ef31047e02fcd1d778c7e4e68fe526c81baa5ec8b5745a26e94d05d13a91fbdd","retrievedAt":"2026-10-30T21:17:56.766Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"6dc41d8d-d245-813b-7afe-22d85381ebb6"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"a5b3776a-7c23-7cf9-2b95-2390a5159dbe"}},"audit":{"status":"uncertain","reviewedAt":"2026-11-22T04:14:53.321Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"a908ba75-1ee3-a83f-34a1-c8c9b596fc97","verdict":"uncertain","findings":[{"dimension":"attribution","status":"uncertain","explanation":"The publisher and speakers are correctly identified. However, the summary sentence \"the approach does not address the scenario of a superhuman model hiding its behavior\" appears in the reading's own voice before Hagendorff is introduced. In the source, this is Hagendorff's observation. The smallest repair is to attribute it to him, for example \"Hagendorff noted that the approach does not address...\".","excerpt":"He also notes that this approach does not address Sutskever’s hypothetical scenario in which a superintelligence hides its true behavior and pretends to be aligned when it isn’t."},{"dimension":"summary","status":"uncertain","explanation":"The central terms \"superalignment\" and \"alignment\" are never explained, and the definitions list is empty. A general reader is not told that alignment means making a model do what you want and avoid what you don't, or that superalignment applies this to hypothetical superhuman models. The reading also does not explain why a stand-in is needed: humans may be unable to score behavior they cannot understand. The smallest repair is one plain-language clause defining alignment and superalignment. Otherwise the summary is faithful and balanced.","excerpt":"Alignment means making sure a model does what you want it to do and does not do what you don’t want it to do. Superalignment applies this idea to superhuman models."},{"dimension":"date","status":"pass","explanation":"December 14, 2023 matches the UTC day of originalPublished (2023-12-14T17:00:18Z). The modification four minutes later is negligible. The Deep Dive sidebar describes later material, such as a story set ten years after AlphaGo, but this is a probable later addition that the reading neither uses nor relies on. The funding is framed as a plan, and the limitations note the retrieval date.","excerpt":null},{"dimension":"support","status":"pass","explanation":"All claims are grounded in the source. These include the GPT-2-supervising-GPT-4 setup, the mixed results (better than GPT-2 on language tasks, worse on chess, and short of what GPT-4 could do alone), the team's conclusion, Hagendorff's critiques, and the planned funding. The question draws on Hagendorff's GPT-3 suggestion and his point about hidden behavior. The conclusion that the results do not establish real-world supervision of superhuman models is appropriately hedged.","excerpt":"But it fell far short of what it could do by itself, he says. They conclude that the approach is promising but needs more work."},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified media candidates were supplied, so a null mediaId is appropriate.","excerpt":null}],"createdAt":"2026-11-22T04:14:53.321Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}