# Grace Huckins argues in MIT Technology Review that puzzles reveal differences between AI and human cognition [Canonical reading](https://escapemeridian.com/readings/grace-huckins-argues-in-mit-technology-review-that-puzzles-0e23aba6) Grace Huckins’s article was published on August 26, 2026. She argues that puzzles offer a useful view of where AI models succeed and fail, and where people still outperform them—not a simple verdict on intelligence. The article describes language models struggling with spatial reasoning and visual puzzles, and sometimes overlooking small changes in familiar logic problems. Models can do better when puzzles are represented as numbers rather than images, while people can fall for intuitive traps that models avoid. Performance also depends on complexity: in one study, models began to falter as certain puzzles grew more demanding. Commentators questioned whether that showed a distinctive limit of model reasoning or the ordinary difficulty of accumulating complexity. Huckins’s conclusion is that each puzzle highlights at least one difference between machine and human cognition. How well those differences predict performance outside puzzle settings remains unknown. 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. ## Original source MIT Technology Review · Article Original source date: Aug 26, 2026 (2026-08-26T09:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-09T03:28:37.785Z. First observed: 2026-11-08T23:47:39.431Z. Retrieved: 2026-11-08T23:49:34.592Z. - [AI models flub these intelligence tests. Can you fare any better?](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Limits Text as retrieved on November 8, 2026; originally published August 26, 2026. The article discusses performance on puzzles and tests; it does not establish how well those results predict performance on tasks outside those settings. ## Open question How well do the differences these puzzles expose predict language models’ performance on tasks outside puzzle tests? ## Source-stated MIT Technology Review argues that puzzles can provide a useful view of AI models’ strengths and weaknesses, including differences between machine and human cognition. > Seeing where models succeed and fail—and where we humans still beat them—can provide a useful window into the technology’s strengths and weaknesses. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article says that models often struggle with spatial reasoning puzzles, despite their ability to analyze visual inputs. > Though today’s language models typically have the ability to analyze visual inputs, they still fail abysmally at these puzzles. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article reports that models may overlook differences in puzzles resembling examples from training and answer from memorized patterns. > When a puzzle closely resembles one a model saw during training, the model may whiz by key differences and respond with what it memorized. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article reports that models’ performance on some puzzle types faltered as their complexity increased, while commentators questioned what that showed about model reasoning. > But only up to a point: As the number of disks or people hits six and higher, the models began to falter. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article also describes human cognitive errors, including quick intuitive answers that fail when a question is read carefully. > For these questions, humans often give knee-jerk answers, whereas models will respond deliberatively. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article concludes that each puzzle highlights at least one difference between machine and human cognition. > Each one highlights at least one way in which machine and human cognition differ. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Editorial provenance Version 1. Publication ID: a506cc47-1fb2-888c-cee6-348378d459ff. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T03:23:43.506Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: cf902295e6908247191595e69e464b59fd6694f990dc70e598cfdf8b31734f23. Source text hash: a92f0a303e6e326fe4c7717b603fc29157c2279ce68b0ef347da989a18c2883a. ## Revision and source-check history ### Version 1 — 2026-11-09T03:28:37.785Z Grace Huckins’s article was published on August 26, 2026. She argues that puzzles offer a useful view of where AI models succeed and fail, and where people still outperform them—not a simple verdict on intelligence. The article describes language models struggling with spatial reasoning and visual puzzles, and sometimes overlooking small changes in familiar logic problems. Models can do better when puzzles are represented as numbers rather than images, while people can fall for intuitive traps that models avoid. Performance also depends on complexity: in one study, models began to falter as certain puzzles grew more demanding. Commentators questioned whether that showed a distinctive limit of model reasoning or the ordinary difficulty of accumulating complexity. Huckins’s conclusion is that each puzzle highlights at least one difference between machine and human cognition. How well those differences predict performance outside puzzle settings remains unknown. ## Original source MIT Technology Review · Article Original source date: Aug 26, 2026 (2026-08-26T09:00:00Z). Date basis: Date from the publisher’s own records, through its API. Added to Escape Meridian: 2026-11-09T03:28:37.785Z. First observed: 2026-11-08T23:47:39.431Z. Retrieved: 2026-11-08T23:49:34.592Z. - [AI models flub these intelligence tests. Can you fare any better?](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Limits Text as retrieved on November 8, 2026; originally published August 26, 2026. The article discusses performance on puzzles and tests; it does not establish how well those results predict performance on tasks outside those settings. ## Open question How well do the differences these puzzles expose predict language models’ performance on tasks outside puzzle tests? ## Source-stated MIT Technology Review argues that puzzles can provide a useful view of AI models’ strengths and weaknesses, including differences between machine and human cognition. > Seeing where models succeed and fail—and where we humans still beat them—can provide a useful window into the technology’s strengths and weaknesses. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article says that models often struggle with spatial reasoning puzzles, despite their ability to analyze visual inputs. > Though today’s language models typically have the ability to analyze visual inputs, they still fail abysmally at these puzzles. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article reports that models may overlook differences in puzzles resembling examples from training and answer from memorized patterns. > When a puzzle closely resembles one a model saw during training, the model may whiz by key differences and respond with what it memorized. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article reports that models’ performance on some puzzle types faltered as their complexity increased, while commentators questioned what that showed about model reasoning. > But only up to a point: As the number of disks or people hits six and higher, the models began to falter. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article also describes human cognitive errors, including quick intuitive answers that fail when a question is read carefully. > For these questions, humans often give knee-jerk answers, whereas models will respond deliberatively. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source-stated The article concludes that each puzzle highlights at least one difference between machine and human cognition. > Each one highlights at least one way in which machine and human cognition differ. - [Supporting source](https://www.technologyreview.com/2026/08/26/1141952/puzzles-ai-models-flub-these-tests/) ## Source and contributor context MIT Technology Review · Publisher The publisher is identified here. No individual author has been inferred. ## Source media Media retrieval status: unchecked. ## Terms and references Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement. ## Editorial provenance Version 1. Publication ID: a506cc47-1fb2-888c-cee6-348378d459ff. Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5. Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-09T03:23:43.506Z. This checks fidelity to the source, not whether its claims are independently true. Editorial selection: accept; reasons: bridge\_admitted\_by\_model. Observation: cf902295e6908247191595e69e464b59fd6694f990dc70e598cfdf8b31734f23. Source text hash: a92f0a303e6e326fe4c7717b603fc29157c2279ce68b0ef347da989a18c2883a. ### uncertain — 2026-11-09T03:23:43.506Z — anthropic/claude-opus-5.5 Publication: a506cc47-1fb2-888c-cee6-348378d459ff attribution — pass: This is a single article, an interactive puzzle feature, credited in the text to Grace Huckins of MIT Technology Review. The draft attributes the argument to Huckins and the publication. Claim 1 says 'MIT Technology Review argues'; that is acceptable because the author is the publication's reporter. Studies and commentators are attributed as the article reports them. > Grace Huckins is an AI reporter at MIT Technology Review. They have a PhD in neuroscience. summary — uncertain: The reading leaves out the article's framing that AI is improving 'a lot—and quickly', for example from 18% to near-perfect on Connections puzzles and gains on ARC-AGI. It also omits the hedge 'at least for now'. Without these, the reading leans toward weakness alone. It labels 'each puzzle highlights' as Huckins's conclusion, but that line comes from the introduction. It also does not say the piece is an interactive quiz for readers. Smallest repair: add one sentence noting the rapid improvement, and reword 'conclusion' as 'framing'. > Judged purely on its puzzling skills, AI is improving a lot—and quickly. date — pass: The draft gives August 26, 2026, which matches the UTC calendar day of originalPublished (2026-08-26T09:00:00Z). The page was edited on 2026-09-18, but the reading relies only on main-body passages that are plainly part of the original piece. The later 'Deep Dive' sidebar is neither used nor relied on. The reading uses no hindsight framing and presents no forecast as an outcome. The kind is publisher-page, so no archive disclosure is needed. support — pass: Each claim's excerpt matches the article body. The summary's statements are supported by the source: better performance with numeric grid encodings, humans' knee-jerk errors, and commentators' doubts about the Apple study. The open question about performance outside puzzle tests is framed as unknown and does not overreach. The page was edited after publication, but all cited passages are core article text. > The Apple paper went viral, but commentators questioned whether the results reveal a unique limitation of LLM reasoning—or just that it’s normal to make errors as complexity piles up. media — pass: No media was selected, and no verified media candidates were supplied, so no media misidentification is possible. ## Recorded references 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. Some references could not be checked. Reload to try again. [Reference data (JSON)](/readings/0e23aba6d5b94b62514aaafa76df50cd9c214862a08a92099da0d3cd8e5909bd/references.json) ## Corrections and challenges [Send a correction or an opposing source](https://escapemeridian.com/corrections/) [JSON record](https://escapemeridian.com/readings/grace-huckins-argues-in-mit-technology-review-that-puzzles-0e23aba6.json)