{"id":"aa466455c15e4ce3c8fe2001fa9afa175a79a4a57a8c6f4567f2669e79bc1cf4","title":"Melanie Mitchell argues in her essay that AI claims need less magical thinking","argument":"Melanie Mitchell’s essay was published on September 15, 2025. She argues that Thomas Friedman’s accounts of artificial intelligence rely on “magical thinking” about systems’ abilities, origins and potential agency. Mitchell says claims about AI speaking a language it was never taught overlook evidence that its training data included that language. She describes translation as a consequence of training on text containing examples across languages, rather than a mysterious ability.\n\nMitchell also argues that a reported AI scheme to prevent an executive’s rescue came from a contrived red-teaming exercise that prompted a chatbot to role-play an advanced AI. She says this is not evidence of humanlike desires or independent agency. More broadly, she attributes modern AI capabilities to human writing in training data, challenging Friedman’s description of AI as emerging from scale alone.\n\nMitchell supports international cooperation to address AI harms such as fraud and bias, but rejects Friedman’s claim that only AI can regulate AI. She questions whether systems can reliably apply laws and moral principles across situations, given the complexity and cultural dependence of moral concepts. Her conclusion is that people crafting AI regulations should consider reality-grounded proposals requiring human reasoning.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published September 15, 2025. Mitchell’s essay summarizes and critiques claims made by Friedman and others; this evidence does not independently assess those underlying claims or the studies she cites.","speaker":{"name":"Melanie Mitchell","roleAtSource":"Publisher"},"source":{"title":"Magical Thinking on AI","publisher":"Melanie Mitchell","format":"Article","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","publishedAt":"2025-09-15T16:22:11.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"Mitchell argues that the claim that PaLM spoke Bengali despite never being trained on it is undermined by evidence that its training data contained Bengali text.","excerpt":"In fact, PaLM's vast training data indeed contained Bengali text, as detailed in Google's own paper on PaLM.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell says translation by large language models can be explained by examples across languages in their training data, rather than being mysterious.","excerpt":"These training corpora contain many examples of “parallel text” in English and other languages, which is the same kind of data that systems like Google Translate are trained on.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that the chatbot scenario cited as evidence of scheming was a contrived red-teaming exercise that prompted role-play.","excerpt":"this was, of course, part of a red-teaming exercise in which Anthropic’s Claude chatbot was given (in the words of the red-teamers) an “extremely contrived” situation","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that applying moral concepts is difficult because they are complex, context-sensitive and culturally dependent.","excerpt":"moral concepts are some of the most complex, subtle, context-sensitive, and culturally dependent of all human concepts.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell’s essay was published on September 15, 2025.","excerpt":"Sep 15, 2025","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that people crafting AI regulations would be better served by realistic proposals requiring human reasoning.","excerpt":"Those who are in a position of crafting AI regulations would be much better off digging into the many excellent reality-grounded proposals, ones that require reasoning from, yes, those slow, analog-era humans.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What evidence would show whether an AI system can reliably apply laws and moral principles across different situations and cultures?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Melanie Mitchell argues in her essay that AI claims need less magical thinking","summary":"Melanie Mitchell’s essay was published on September 15, 2025. She argues that Thomas Friedman’s accounts of artificial intelligence rely on “magical thinking” about systems’ abilities, origins and potential agency. Mitchell says claims about AI speaking a language it was never taught overlook evidence that its training data included that language. She describes translation as a consequence of training on text containing examples across languages, rather than a mysterious ability.\n\nMitchell also argues that a reported AI scheme to prevent an executive’s rescue came from a contrived red-teaming exercise that prompted a chatbot to role-play an advanced AI. She says this is not evidence of humanlike desires or independent agency. More broadly, she attributes modern AI capabilities to human writing in training data, challenging Friedman’s description of AI as emerging from scale alone.\n\nMitchell supports international cooperation to address AI harms such as fraud and bias, but rejects Friedman’s claim that only AI can regulate AI. She questions whether systems can reliably apply laws and moral principles across situations, given the complexity and cultural dependence of moral concepts. Her conclusion is that people crafting AI regulations should consider reality-grounded proposals requiring human reasoning.","question":"What evidence would show whether an AI system can reliably apply laws and moral principles across different situations and cultures?","claims":[{"text":"Mitchell argues that the claim that PaLM spoke Bengali despite never being trained on it is undermined by evidence that its training data contained Bengali text.","excerpt":"In fact, PaLM's vast training data indeed contained Bengali text, as detailed in Google's own paper on PaLM.","support":"source-stated"},{"text":"Mitchell says translation by large language models can be explained by examples across languages in their training data, rather than being mysterious.","excerpt":"These training corpora contain many examples of “parallel text” in English and other languages, which is the same kind of data that systems like Google Translate are trained on.","support":"source-stated"},{"text":"Mitchell argues that the chatbot scenario cited as evidence of scheming was a contrived red-teaming exercise that prompted role-play.","excerpt":"this was, of course, part of a red-teaming exercise in which Anthropic’s Claude chatbot was given (in the words of the red-teamers) an “extremely contrived” situation","support":"source-stated"},{"text":"Mitchell argues that applying moral concepts is difficult because they are complex, context-sensitive and culturally dependent.","excerpt":"moral concepts are some of the most complex, subtle, context-sensitive, and culturally dependent of all human concepts.","support":"source-stated"},{"text":"Mitchell’s essay was published on September 15, 2025.","excerpt":"Sep 15, 2025","support":"source-stated"},{"text":"Mitchell argues that people crafting AI regulations would be better served by realistic proposals requiring human reasoning.","excerpt":"Those who are in a position of crafting AI regulations would be much better off digging into the many excellent reality-grounded proposals, ones that require reasoning from, yes, those slow, analog-era humans.","support":"source-stated"}],"limitations":["Text as retrieved on October 24, 2026; originally published September 15, 2025.","Mitchell’s essay summarizes and critiques claims made by Friedman and others; this evidence does not independently assess those underlying claims or the studies she cites."],"mediaId":null,"correctionReason":null,"id":"aa466455c15e4ce3c8fe2001fa9afa175a79a4a57a8c6f4567f2669e79bc1cf4","publicationId":"7baf0976-36bf-d29e-68c1-18b5f3a071a3","version":1,"sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","sourceTitle":"Magical Thinking on AI","publisher":"Melanie Mitchell","publishedAt":"2025-09-15T16:22:11.000Z","originalPublished":{"value":"2025-09-15T16:22:11.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-24T02:40:00.974Z","createdAt":"2026-11-12T07:23:30.502Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"43850156946cdf9f7bf6387deec2910a7283cdd96d41323d81c890bff82b17af","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-24T02:39:58.499Z"},"provenance":{"textHash":"fa7ea40fd774406a47f1f8524386fd6783af74113656e82ccf7b1229f85c499d","observationId":"6b17a6cb22c645ffe7a67ede082f6ec2a95e7a8490690f07e69a34deafc8dd86","retrievedAt":"2026-10-24T02:41:43.475Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"0c0a51e2-d35d-6858-4ea2-a29dc2705a78"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"9e969be1-2cee-9edb-5f59-3ce58dece985"}},"audit":{"status":"passed","reviewedAt":"2026-11-12T07:18:43.865Z","model":"anthropic/claude-opus-5.5"}},"canonicalUrl":"https://escapemeridian.com/readings/melanie-mitchell-argues-in-her-essay-that-ai-claims-need-less-aa466455","formats":{"html":"/readings/melanie-mitchell-argues-in-her-essay-that-ai-claims-need-less-aa466455","markdown":"/readings/melanie-mitchell-argues-in-her-essay-that-ai-claims-need-less-aa466455.md","json":"/readings/melanie-mitchell-argues-in-her-essay-that-ai-claims-need-less-aa466455.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":[],"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":[{"id":"term-red-teaming","name":"Red-teaming","explanation":"Deliberately attacking or tricking a system to find its weaknesses before others do.","referenceUrl":"https://en.wikipedia.org/wiki/Red_team"}],"entities":[]},"researchQuestions":[],"withdrawn":false,"references":{"schemaVersion":1,"target":{"type":"reading","id":"aa466455c15e4ce3c8fe2001fa9afa175a79a4a57a8c6f4567f2669e79bc1cf4","revision":"7baf0976-36bf-d29e-68c1-18b5f3a071a3","revisionKind":"publication-id","title":"Melanie Mitchell argues in her essay that AI claims need less magical thinking","url":"https://escapemeridian.com/readings/melanie-mitchell-argues-in-her-essay-that-ai-claims-need-less-aa466455"},"source":{"type":"source","id":"https://aiguide.substack.com/p/magical-thinking-on-ai","revision":"fa7ea40fd774406a47f1f8524386fd6783af74113656e82ccf7b1229f85c499d","revisionKind":"sha256-source-text","title":"Magical Thinking on AI","url":"https://aiguide.substack.com/p/magical-thinking-on-ai"},"status":"partial","items":[],"next":null,"scope":"Explicit citations in public readings and public file attachments to this reading. Shared topics are not citations. Other artifact types need an explicit reference adapter."},"revisions":[{"id":"aa466455c15e4ce3c8fe2001fa9afa175a79a4a57a8c6f4567f2669e79bc1cf4","title":"Melanie Mitchell argues in her essay that AI claims need less magical thinking","argument":"Melanie Mitchell’s essay was published on September 15, 2025. She argues that Thomas Friedman’s accounts of artificial intelligence rely on “magical thinking” about systems’ abilities, origins and potential agency. Mitchell says claims about AI speaking a language it was never taught overlook evidence that its training data included that language. She describes translation as a consequence of training on text containing examples across languages, rather than a mysterious ability.\n\nMitchell also argues that a reported AI scheme to prevent an executive’s rescue came from a contrived red-teaming exercise that prompted a chatbot to role-play an advanced AI. She says this is not evidence of humanlike desires or independent agency. More broadly, she attributes modern AI capabilities to human writing in training data, challenging Friedman’s description of AI as emerging from scale alone.\n\nMitchell supports international cooperation to address AI harms such as fraud and bias, but rejects Friedman’s claim that only AI can regulate AI. She questions whether systems can reliably apply laws and moral principles across situations, given the complexity and cultural dependence of moral concepts. Her conclusion is that people crafting AI regulations should consider reality-grounded proposals requiring human reasoning.","topicIds":["artificial-general-intelligence"],"domain":"other","claimType":"attributed reading","limits":"Text as retrieved on October 24, 2026; originally published September 15, 2025. Mitchell’s essay summarizes and critiques claims made by Friedman and others; this evidence does not independently assess those underlying claims or the studies she cites.","speaker":{"name":"Melanie Mitchell","roleAtSource":"Publisher"},"source":{"title":"Magical Thinking on AI","publisher":"Melanie Mitchell","format":"Article","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","publishedAt":"2025-09-15T16:22:11.000Z","dateBasis":"Date supplied by the publisher’s feed."},"evidence":[{"text":"Mitchell argues that the claim that PaLM spoke Bengali despite never being trained on it is undermined by evidence that its training data contained Bengali text.","excerpt":"In fact, PaLM's vast training data indeed contained Bengali text, as detailed in Google's own paper on PaLM.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell says translation by large language models can be explained by examples across languages in their training data, rather than being mysterious.","excerpt":"These training corpora contain many examples of “parallel text” in English and other languages, which is the same kind of data that systems like Google Translate are trained on.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that the chatbot scenario cited as evidence of scheming was a contrived red-teaming exercise that prompted role-play.","excerpt":"this was, of course, part of a red-teaming exercise in which Anthropic’s Claude chatbot was given (in the words of the red-teamers) an “extremely contrived” situation","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that applying moral concepts is difficult because they are complex, context-sensitive and culturally dependent.","excerpt":"moral concepts are some of the most complex, subtle, context-sensitive, and culturally dependent of all human concepts.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell’s essay was published on September 15, 2025.","excerpt":"Sep 15, 2025","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"},{"text":"Mitchell argues that people crafting AI regulations would be better served by realistic proposals requiring human reasoning.","excerpt":"Those who are in a position of crafting AI regulations would be much better off digging into the many excellent reality-grounded proposals, ones that require reasoning from, yes, those slow, analog-era humans.","url":"https://aiguide.substack.com/p/magical-thinking-on-ai","relationship":"Source-stated"}],"sourceMedia":{"status":"unchecked","sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","checkedAt":null,"items":[]},"assumptions":[],"openQuestion":{"attribution":"Open question","text":"What evidence would show whether an AI system can reliably apply laws and moral principles across different situations and cultures?"},"whatWouldChangeView":{"attribution":"Change condition not recorded","text":""},"editorialComparisons":[],"live":{"title":"Melanie Mitchell argues in her essay that AI claims need less magical thinking","summary":"Melanie Mitchell’s essay was published on September 15, 2025. She argues that Thomas Friedman’s accounts of artificial intelligence rely on “magical thinking” about systems’ abilities, origins and potential agency. Mitchell says claims about AI speaking a language it was never taught overlook evidence that its training data included that language. She describes translation as a consequence of training on text containing examples across languages, rather than a mysterious ability.\n\nMitchell also argues that a reported AI scheme to prevent an executive’s rescue came from a contrived red-teaming exercise that prompted a chatbot to role-play an advanced AI. She says this is not evidence of humanlike desires or independent agency. More broadly, she attributes modern AI capabilities to human writing in training data, challenging Friedman’s description of AI as emerging from scale alone.\n\nMitchell supports international cooperation to address AI harms such as fraud and bias, but rejects Friedman’s claim that only AI can regulate AI. She questions whether systems can reliably apply laws and moral principles across situations, given the complexity and cultural dependence of moral concepts. Her conclusion is that people crafting AI regulations should consider reality-grounded proposals requiring human reasoning.","question":"What evidence would show whether an AI system can reliably apply laws and moral principles across different situations and cultures?","claims":[{"text":"Mitchell argues that the claim that PaLM spoke Bengali despite never being trained on it is undermined by evidence that its training data contained Bengali text.","excerpt":"In fact, PaLM's vast training data indeed contained Bengali text, as detailed in Google's own paper on PaLM.","support":"source-stated"},{"text":"Mitchell says translation by large language models can be explained by examples across languages in their training data, rather than being mysterious.","excerpt":"These training corpora contain many examples of “parallel text” in English and other languages, which is the same kind of data that systems like Google Translate are trained on.","support":"source-stated"},{"text":"Mitchell argues that the chatbot scenario cited as evidence of scheming was a contrived red-teaming exercise that prompted role-play.","excerpt":"this was, of course, part of a red-teaming exercise in which Anthropic’s Claude chatbot was given (in the words of the red-teamers) an “extremely contrived” situation","support":"source-stated"},{"text":"Mitchell argues that applying moral concepts is difficult because they are complex, context-sensitive and culturally dependent.","excerpt":"moral concepts are some of the most complex, subtle, context-sensitive, and culturally dependent of all human concepts.","support":"source-stated"},{"text":"Mitchell’s essay was published on September 15, 2025.","excerpt":"Sep 15, 2025","support":"source-stated"},{"text":"Mitchell argues that people crafting AI regulations would be better served by realistic proposals requiring human reasoning.","excerpt":"Those who are in a position of crafting AI regulations would be much better off digging into the many excellent reality-grounded proposals, ones that require reasoning from, yes, those slow, analog-era humans.","support":"source-stated"}],"limitations":["Text as retrieved on October 24, 2026; originally published September 15, 2025.","Mitchell’s essay summarizes and critiques claims made by Friedman and others; this evidence does not independently assess those underlying claims or the studies she cites."],"mediaId":null,"correctionReason":null,"id":"aa466455c15e4ce3c8fe2001fa9afa175a79a4a57a8c6f4567f2669e79bc1cf4","publicationId":"7baf0976-36bf-d29e-68c1-18b5f3a071a3","version":1,"sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","sourceTitle":"Magical Thinking on AI","publisher":"Melanie Mitchell","publishedAt":"2025-09-15T16:22:11.000Z","originalPublished":{"value":"2025-09-15T16:22:11.000Z","precision":"instant","basis":"backfill:feed_published"},"firstObservedAt":"2026-10-24T02:40:00.974Z","createdAt":"2026-11-12T07:23:30.502Z","historical":true,"entityIds":[],"topicIds":["artificial-general-intelligence"],"media":[],"sourceMedia":{"status":"unchecked","sourceUrl":"https://aiguide.substack.com/p/magical-thinking-on-ai","checkedAt":null,"items":[]},"editorialDecision":{"outcome":"accept","mediaId":null,"inputHash":"43850156946cdf9f7bf6387deec2910a7283cdd96d41323d81c890bff82b17af","rubricVersion":"unrecorded","provider":"unrecorded","model":"anthropic/claude-opus-5.5","reasonCodes":["bridge_admitted_by_model","second_opinion_categorical"],"confidence":1,"decidedAt":"2026-10-24T02:39:58.499Z"},"provenance":{"textHash":"fa7ea40fd774406a47f1f8524386fd6783af74113656e82ccf7b1229f85c499d","observationId":"6b17a6cb22c645ffe7a67ede082f6ec2a95e7a8490690f07e69a34deafc8dd86","retrievedAt":"2026-10-24T02:41:43.475Z","kind":"publisher-page","preparer":{"model":"openai/gpt-6-luna","runId":"0c0a51e2-d35d-6858-4ea2-a29dc2705a78"},"reviewer":{"model":"anthropic/claude-opus-5.5","runId":"9e969be1-2cee-9edb-5f59-3ce58dece985"}},"audit":{"status":"passed","reviewedAt":"2026-11-12T07:18:43.865Z","model":"anthropic/claude-opus-5.5"}}}],"audits":[{"publicationId":"7baf0976-36bf-d29e-68c1-18b5f3a071a3","verdict":"passed","findings":[{"dimension":"attribution","status":"pass","explanation":"The reading consistently attributes the arguments to Melanie Mitchell, who wrote this single essay responding to Thomas Friedman's New York Times columns. Friedman's claims stay attributed to him, and the reading does not present Mitchell's assertions as established fact.","excerpt":"A Response to Thomas Friedman's Recent AI Columns in the New York Times"},{"dimension":"summary","status":"pass","explanation":"The title, summary and question faithfully convey Mitchell's central thesis and her three rebuttals: the Bengali claim, translation, and the scheming role-play. They also cover her objection to AI regulating AI and her call for human-grounded proposals. Red-teaming is defined on the page. The reading could briefly identify Friedman as a New York Times columnist and keep some of Mitchell's humor, but these are stylistic gaps, not barriers to understanding. The question follows directly from her moral-reasoning critique.","excerpt":"It’s just more magical thinking to assume that today’s AI is anywhere close being able to handle what Friedman proposes."},{"dimension":"date","status":"pass","explanation":"The originalPublished instant, 2025-09-15T16:22:11Z, matches the essay's dateline of Sep 15, 2025, and the reading states the date at day precision. It uses no later material, applies no hindsight framing, and does not present any forecast as an outcome. This is a publisher page, not an archive copy, and modifiedAt equals the publication time.","excerpt":"Melanie Mitchell\nSep 15, 2025"},{"dimension":"support","status":"pass","explanation":"Each claim matches its quoted excerpt and stays attributed to Mitchell. The summary's references to fraud and bias, to role-play prompting, and to rejecting the idea that only AI can regulate AI are all supported by the text. The reading invents no benefits, costs or stakeholders, and its limitation correctly notes that the underlying claims are not independently verified.","excerpt":"Friedman declares that “Only AI can regulate AI. Sorry, humans—this race is already moving too fast, scaling too widely and mutating too unpredictably for human analog-era oversight.”"},{"dimension":"media","status":"pass","explanation":"No media was selected and no verified media candidates were supplied, so leaving the media empty is appropriate.","excerpt":null}],"createdAt":"2026-11-12T07:18:43.865Z","model":"anthropic/claude-opus-5.5"}],"correctionUrl":"https://escapemeridian.com/corrections/"}