--- title: System helps humans predict when self-driving cars will... description: Researchers from MIT and Motional developed CW-Net, a Concept-Wrapper Network that translates the opaque internal reasoning of self-driving car planners into... canonical: https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: System helps humans predict when self-driving cars will make mistakes | daily.dev og:description: Researchers from MIT and Motional developed CW-Net, a Concept-Wrapper Network that translates the opaque internal reasoning of self-driving car planners into... og:url: https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf og:image: https://api.daily.dev/og/posts/ZXduamwwf.png og:image:alt: System helps humans predict when self-driving cars will make mistakes og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # System helps humans predict when self-driving cars will make mistakes **[MIT News](https://daily.dev/sources/mit)** · 6 min read · 1 upvotes · 1 comments ## Summary Researchers from MIT and Motional developed CW-Net, a Concept-Wrapper Network that translates the opaque internal reasoning of self-driving car planners into human-understandable concepts like 'approaching stopped vehicle' or 'close to cyclist.' The module plugs into an existing machine-learning planner architecture, forcing the final decision stage to use these concepts, which ensures the explanations are causally faithful rather than misleading. In track tests with a Motional robotaxi and safety driver, CW-Net revealed that a vehicle stopped near a cyclist not because it detected the cyclist, but because emergency braking kicked in after a misconfigured detection nearly caused a collision. Both real-world track tests and larger online simulations using Las Vegas driving data showed CW-Net significantly improved people's ability to predict vehicle behavior in surprising situations, without degrading driving performance. The work, published in Nature, is aimed at improving safety, transparency, and engineering diagnostics for autonomous vehicle AI systems. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### How does CW-Net make self-driving car AI decisions more understandable to humans? CW-Net is a concept classifier plugged into the middle of an autonomous vehicle's machine-learning planner that translates opaque internal reasoning into human-readable concepts like 'approaching stopped vehicle' or 'close to cyclist.' It forces the final decision layer to use these concepts, ensuring explanations causally reflect the model's actual reasoning rather than being a misleading afterthought, without changing driving performance. _Developers exploring AI interpretability techniques can follow research like this on daily.dev._ ### What real-world failure did CW-Net reveal in a Motional robotaxi's cyclist detection? In track testing, a Motional robotaxi consistently stopped near cyclists, and the safety driver assumed it correctly detected them. CW-Net explanations instead showed the model had not properly detected the cyclist and had chosen a trajectory that would have caused a collision; the vehicle stopped only because emergency braking triggered when it got too close. _Engineers debugging safety-critical AI systems can track findings like this on daily.dev._ ## Community discussion Top comments from developers on daily.dev. **@empera0** · 0 upvotes > I don't know how this is gonna be the solution. Like rn with with vibe-coding even that AI's does what he told sometimes and can smack everything inside the code. and like my first question is gonna be - "What if AI starts doing miscalculation about what it says and what it does?" > > In a example There is a crossroad and right side of the car there is a bicycle okay. AI says "there is a bicycle in my right i am not turning right i stop" then just turns right? What are they gonna do about that beacuse this is the real problem about the AI rn. I can understand getting perfection in a highscore... ## Similar posts on daily.dev - [Waymo says it built a better benchmark for comparing robotaxis to humans](https://daily.dev/posts/waymo-says-it-built-a-better-benchmark-for-comparing-robotaxis-to-humans-yjbulwjno) · TechCrunch · 0 upvotes · 0 comments - [The Waymo World Model: A New Frontier For Autonomous Driving Simulation](https://daily.dev/posts/the-waymo-world-model-a-new-frontier-for-autonomous-driving-simulation-5jsu87osj) · Hacker News · 0 upvotes · 0 comments - [A Korean AI model scores every possible driving path for safety before the car moves. CVPR called it a highlight.](https://daily.dev/posts/a-korean-ai-model-scores-every-possible-driving-path-for-safety-before-the-car-moves-cvpr-called-it-a8ecsusr4) · The Next Web · 0 upvotes · 0 comments - [Self-driving cars, drones hijacked by custom road signs](https://daily.dev/posts/self-driving-cars-drones-hijacked-by-custom-road-signs-fehi3msia) · The Register · 1 upvotes · 0 comments --- Tags: [#deep-learning](https://daily.dev/tags/deep-learning), [#mit](https://daily.dev/tags/mit), [#ai-safety](https://daily.dev/tags/ai-safety), [#explainable-ai](https://daily.dev/tags/explainable-ai) [View this post on daily.dev](https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"System helps humans predict when self-driving cars will make mistakes","url":"https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf"},"datePublished":"2026-09-02T15:21:28.455Z","dateModified":"2026-09-14T09:03:59.840Z","description":"Researchers from MIT and Motional developed CW-Net, a Concept-Wrapper Network that translates the opaque internal reasoning of self-driving car planners into...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c8efe29cb17e4db03df11996d4d42e4b?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/c8efe29cb17e4db03df11996d4d42e4b?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"MIT News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"MIT News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/881e20a40aa44b12ba4bd27f5c14289d","url":"https://daily.dev/sources/mit"},"commentCount":1,"discussionUrl":"https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":1}],"keywords":"deep-learning,mit,ai-safety,explainable-ai","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"MIT News","item":"https://daily.dev/sources/mit"},{"@type":"ListItem","position":3,"name":"System helps humans predict when self-driving cars will make mistakes"}]} {"@context":"https://schema.org","@type":"WebPage","@id":"https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf","comment":[{"@type":"Comment","text":"I don’t know how this is gonna be the solution. Like rn with with vibe-coding even that AI’s does what he told sometimes and can smack everything inside the code. and like my first question is gonna be - “What if AI starts doing miscalculation about what it says and what it does?”\nIn a example There is a crossroad and right side of the car there is a bicycle okay. AI says “there is a bicycle in my right i am not turning right i stop” then just turns right? What are they gonna do about that beacuse this is the real problem about the AI rn. I can understand getting perfection in a highscore with Human answers but this is not gonna change and there is gonna be mistakes happen about that. It’s about the lines around neurons what we gave to it. There is a good side of it and bad side of it and the one of choices becames the part of the correlation around the good ones then it is gonna make mistake again.","datePublished":"2026-09-02T15:46:59.435Z","url":"https://daily.dev/posts/ZXduamwwf#c-CLJncNyLq","author":{"@type":"Person","name":"Oğuz Görencan","url":"https://daily.dev/empera0","image":"https://avatars.githubusercontent.com/u/63954402?v=4"}}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/system-helps-humans-predict-when-self-driving-cars-will-make-mistakes-zxduamwwf#faq","mainEntity":[{"@type":"Question","name":"How does CW-Net make self-driving car AI decisions more understandable to humans?","acceptedAnswer":{"@type":"Answer","text":"CW-Net is a concept classifier plugged into the middle of an autonomous vehicle's machine-learning planner that translates opaque internal reasoning into human-readable concepts like 'approaching stopped vehicle' or 'close to cyclist.' It forces the final decision layer to use these concepts, ensuring explanations causally reflect the model's actual reasoning rather than being a misleading afterthought, without changing driving performance. Developers exploring AI interpretability techniques can follow research like this on daily.dev."}},{"@type":"Question","name":"What real-world failure did CW-Net reveal in a Motional robotaxi's cyclist detection?","acceptedAnswer":{"@type":"Answer","text":"In track testing, a Motional robotaxi consistently stopped near cyclists, and the safety driver assumed it correctly detected them. CW-Net explanations instead showed the model had not properly detected the cyclist and had chosen a trajectory that would have caused a collision; the vehicle stopped only because emergency braking triggered when it got too close. Engineers debugging safety-critical AI systems can track findings like this on daily.dev."}}]} ```