Browse claims and the cases made for and against them. Find one to support, dispute, or add context to.
28 matching claims · Recently updated first
The ten-to-one comparison rests on a 2023 estimate of about 3 Wh per ChatGPT request. A 2025 analysis puts a typical GPT-4o request near 0.3 Wh, OpenAI's chief executive gives 0.34 Wh, and Google measures 0.24 Wh for a median Gemini text prompt. Those figures are close to the long-quoted number for a search. Estimates depend on model, response length and what is counted.
1 supporting · 0 disputing
GPT-4 scored strongly against physicians on written diagnostic cases, and Google's AMIE matched or exceeded clinicians in simulated consultations. A pooled analysis of 83 studies found no significant difference from non-specialist physicians, though AI performed below experts. The evidence comes from test cases and simulated patients, not routine care of real patients.
1 supporting · 0 disputing
AlphaFold predicts many protein structures accurately, but its predictions still need experimental checking, and for proteins that can adopt two folds it usually predicts only one. The Nobel citation credits structure prediction, and the successor model is still extending to complexes. Predicting a structure is not the same as explaining how proteins fold, move and interact.
1 supporting · 0 disputing
Researchers have extracted memorised training images from diffusion models and shown near-copies appearing in output, especially for images repeated in the training data. The US Copyright Office has reviewed the same evidence. The claim is that reproduction can happen, not that most outputs are copies or that any particular output infringes copyright.
1 supporting · 0 disputing
The US Copyright Office, a federal district court and the DC Circuit have each held that copyright requires a human author, and registration was refused for a work presented as generated autonomously by a machine. A human's selection, arrangement or modification of AI material can still be protected, as in a 2023 decision on a comic with AI-generated images. UK law treats computer-generated works differently.
1 supporting · 0 disputing
Studies of generated bibliographies, legal questions, commercial legal research tools and systematic-review searches all found language models producing references that do not exist. Rates differ widely by model and task, and one comparison found negligible fabrication from retrieval-based tools. Older measured rates should not be assumed for every current system.
1 supporting · 0 disputing
Tests of detection tools found poor reliability, frequent mislabelling of writing by non-native English speakers, and easy evasion through paraphrasing. OpenAI withdrew its own classifier for low accuracy. A claim of reliability needs a named detector, version, population and error rate.
1 supporting · 0 disputing
A Swedish randomised trial, a large German implementation study and a separate Swedish prospective study each found more cancers detected when radiologists read with AI support than under standard double reading. Follow-up of the Swedish trial found higher sensitivity with a non-inferior interval-cancer rate. Higher detection is not evidence of lower mortality.
1 supporting · 0 disputing
Examines the reported percentile of GPT-4’s simulated US Uniform Bar Examination performance. Its passing score is different from its percentile, which depends on the comparison cohort and how essays are scored. Reanalysis challenges the widely repeated 90th-percentile figure.
0 supporting · 0 disputing · 1 adding context
Examines whether novices gain more than experienced workers from generative AI within the same job or task. Evidence from customer support, writing and consulting often shows larger gains for initially weaker performers; this is not a rule for every profession.
1 supporting · 0 disputing
Tests whether access to AI coding assistants shortens development tasks. Controlled exercises and workplace studies show gains, while experienced developers working in familiar repositories have sometimes taken longer. Task complexity, tool generation and familiarity matter.
1 supporting · 0 disputing
Compares chatbot diagnostic performance with physicians. Strong scores on written vignettes and simulated consultations coexist with failures in more realistic clinical tasks. Model-only accuracy, clinician-assisted accuracy and safe patient care are distinct outcomes.
1 supporting · 0 disputing
Checks whether private browsing makes online activity invisible to websites and internet providers. Browser documentation explicitly limits the feature mainly to local history and session storage. It does not provide network anonymity or prevent sites from seeing a visit.
0 supporting · 4 disputing
The scientific community broadly agrees that social media can both benefit and harm children and adolescents. As stated in the APA's Health Advisory on Social Media Use in Adolescence, using social media is not inherently beneficial or harmful to young people. Adolescents' online lives both reflect and shape their offline lives. In most cases, social media effects depend on adolescents' personal and psychological characteristics and social circumstances, intersecting with the specific content, features, or functions offered by many social media platforms. In other words, social media effects likely depend on what teens can do and see online, their pre-existing strengths or vulnerabilities, and the contexts in which they grow up. It is agreed that social media bans on their own do not automatically produce relief. Still, they can be part of the equation, as long as they're healthy bans that invest sufficient resources to support a rigorous evaluation.
0 supporting · 1 disputing · 2 adding context
Compares security weaknesses in code produced with and without AI assistance. Some controlled studies find more insecure solutions and misplaced confidence, while another user study finds no substantial increase in severe bugs. Generated-code benchmarks alone cannot establish the human comparison.
2 supporting · 1 disputing
SPAC listings have surged again in 2026 despite the poor performance of many previous deals. Reuters reports that only around 10% of previously completed SPAC deals currently trade above their original $10 price, while more than a third trade below $1.
No cases yet
Autonomous-driving companies increasingly claim their vehicles crash less often than humans. Waymo says its vehicles have substantially fewer serious crashes under comparable conditions, but regulators continue investigating incidents involving driverless systems and competing safety claims remain controversial.
No cases yet
Chinese passenger-car exports jumped roughly 78% year-on-year in August while domestic sales fell almost 24%. The claim asks whether China's export surge reflects highly competitive products or manufacturers being forced overseas by a saturated domestic market.
No cases yet
Huawei reportedly supplies about 52% of Africa's 5G infrastructure. The US is now financing competing telecom infrastructure, arguing that reliance on Chinese equipment creates security risks, while Huawei has consistently denied allegations that its technology enables espionage.
No cases yet
India has become one of Russia's largest oil customers despite Western sanctions imposed after the invasion of Ukraine. The claim examines whether Indian purchases materially sustain Russian revenues or simply redirect oil that would otherwise find other buyers.
No cases yet