Hype Ledger

This ledger tracks AI claims we have checked, who made them, and what the evidence shows. Each entry lists open questions, a plain status, and the date we last checked it. A status can change as more evidence comes in. There are no scores, just these four statuses:

  • Supported: The main sources back up what was claimed.
  • Promising but unverified: There is evidence, but it comes only from the people behind the claim.
  • Needs follow-up: Important details lack confirmation or depend on future events.
  • Did not hold up: The evidence goes against the claim.

Few DoorDash orders come from AI agents

Claim
Outside AI shopping agents have placed only “several hundred” DoorDash orders in total.
Who made it
DoorDash CEO Tony Xu, during the third-quarter earnings call.
Date
Earnings call reported October 30, 2026.
Primary source
TechCrunch.
Independent check
No independent check found. Ars Technica separately reported the same remark from the call. For context, CNBC reported 993 million total orders for the quarter.
Still unproven
The count is a rough company estimate. It is unclear whether use will grow through DoorDash’s live MCP partner connector. Ars says it is less than a month old and serves a small beta group. Xu’s ideas about agents helping with larger purchases or ads remain untested.
Status
Supported
Last checked
2026-11-03

Read our check

A model turns Firefox fixes into attacks

Claim
The freely available GLM 5.3 model made working attacks against every Firefox release from September 17 to October 26. It used published fixes that had not yet reached users. After setup, the cost was about $1,000 for each version.
Who made it
Epoch AI researchers.
Date
October 29, 2026.
Primary source
Epoch AI report.
Independent check
Mozilla agrees the risk is real and has changed how it ships fixes. It moved to weekly builds and keeps regression tests private until fixes reach users. Firefox 159.0 fixes a bug reported by Epoch.
Still unproven
Epoch expects attacks on Firefox 159.0 to be much harder, but that remains to be tested. It is unclear whether an attacker with little skill could repeat the work outside Epoch’s test setting. Setup cost more than $100,000. The method has not yielded a working Chrome attack.
Status
Supported
Last checked
2026-11-03

Read our check

Analysts follow an agent’s random preference

Claim
In a study of 50 trained analysts, participants followed an AI agent’s preference 86% of the time. A hidden coin flip set that preference. Most later used the agent’s reasoning to explain their own choices.
Who made it
Anthropic researchers.
Date
October 27, 2026.
Primary source
Anthropic research post.
Independent check
No independent check found.
Still unproven
The effect in real work is unknown. Anthropic deliberately strengthened the agents’ preferences. The study used one model and one fictional farming task. The group working without an agent was small and had less experience.
Status
Promising but unverified
Last checked
2026-11-03

OpenAI reportedly cuts its funding target

Claim
OpenAI reduced its funding goal from a reported $30 billion to $20 billion. It reportedly has $11 billion committed after some large investors showed limited interest.
Who made it
The Financial Times reported the target change. The Wall Street Journal reported the $11 billion figure. TechCrunch covered both reports.
Date
Financial Times report on October 30, 2026; TechCrunch coverage on October 31, 2026.
Primary source
TechCrunch.
Independent check
No independent check found. The reports use unnamed sources, and OpenAI has not commented.
Still unproven
The amount in the first funding close is unknown. So is whether the reported $1.25 trillion value before new funding will fall again. OpenAI has not confirmed the reports.
Status
Needs follow-up
Last checked
2026-11-03

Altman questions RAM use in home computers

Claim
Sam Altman called it “quite terrible” that scarce DRAM sits in personal computers and is used about five percent of the time. He said tight supplies should go toward fuller use in the cloud.
Who made it
OpenAI CEO Sam Altman, on the Cheeky Pint podcast.
Date
Reported October 30, 2026.
Primary source
The Verge.
Independent check
Tom’s Hardware examined the comment. It noted that the argument benefits OpenAI’s data center business and assumes a particular way to judge memory use. The supply shortage is real.
Still unproven
Altman gave no source for the five percent figure. It is unclear whether constant AI use creates more value than home use. A usage rate alone does not settle the value of privacy, ownership, or offline access.
Status
Needs follow-up
Last checked
2026-11-03

Apple’s CEO backs its AI direction

Claim
Apple is in a good spot on AI. John Ternus said “I really like where we’re at” and that Apple will use AI in its products but does not want to ship “AI for AI’s sake.”
Who made it
Apple CEO John Ternus, in an interview with WIRED’s Steven Levy.
Date
TechCrunch coverage on October 31, 2026.
Primary source
TechCrunch.
Independent check
No independent check found.
Still unproven
The reported interview offers no new product, measure, or deadline for testing the claim. Apple’s AI rollout in China is also pending. The Verge reports an iOS 27.3 beta set for November 4 and a full release planned for December.
Status
Needs follow-up
Last checked
2026-11-03

Claude checks its own chats for possible harm

Claim
Having Claude assess a chat just after replying catches 78% of chats that a stronger classifier labels as possibly harmful. It catches 94% of high-priority cases, with a 1% false-positive rate and less than 2% added computing work.
Who made it
Anthropic’s Safeguards team.
Date
October 16, 2026.
Primary source
Anthropic Safeguards.
Independent check
No independent check found.
Still unproven
Another automated classifier supplies the comparison labels. The rates do not measure confirmed harm in the real world. The post does not show how often closer review overturns a flag. The method picks chats for later review. It does not block replies.
Status
Promising but unverified
Last checked
2026-11-03

AMD shows lower token costs in a benchmark

Claim
At 100 tokens per second for each user, serving Kimi K3 costs $0.077 per million tokens on AMD MI355X chips. The reported cost on Nvidia B200 chips is $0.179.
Who made it
InferenceX, SemiAnalysis’s benchmark dashboard.
Date
The live page has no publication date. The figures were confirmed there on November 3, 2026.
Primary source
InferenceX comparison.
Independent check
No independent check found.
Still unproven
The costs are filled in between measured benchmark points. They use hourly chip prices from SemiAnalysis’s own surveys. The default test replays coding agent sessions and often reuses cached data. Other kinds of work may cost differently. No outside team has reproduced the result.
Status
Promising but unverified
Last checked
2026-11-03

Reddit links lost search visits to AI

Claim
AI is causing a sizable fall in visits to Reddit from Google search.
Who made it
Reddit CEO Steve Huffman, speaking with CNBC after third-quarter results.
Date
October 30, 2026.
Primary source
CNBC.
Independent check
No independent check found. Reddit’s shareholder letter records fewer search referrals, but those are company figures.
Still unproven
The share of lost visits caused by AI, rather than other search changes, is unknown. The stock fell 29.22% in one day, but the role of traffic concerns versus a weak fourth-quarter sales forecast is unclear. A planned test requiring logins has yet to show whether it brings users back.
Status
Needs follow-up
Last checked
2026-11-03

Models struggle on GDPx expert tasks

Claim
The best model reaches only 12% on GDPx, SemiAnalysis’s new test of difficult expert work.
Who made it
SemiAnalysis.
Date
November 2, 2026.
Primary source
SemiAnalysis.
Independent check
No independent check found. Neural Brew reported Claude Opus 5.5 leading at 12% and said the results lack outside reproduction.
Still unproven
The selected fields may not represent expert work as a whole. SemiAnalysis says safety classifiers lower some results, but their effect on the ranking is unclear. Most tasks are private. Firms that build training environments for AI labs also built these tasks, leaving questions about whether outsiders can repeat the tests.
Status
Promising but unverified
Last checked
2026-11-03

Hype Over Time

Older promises, checked years later. See the full Hype Over Time scorecard.

Tesla robotaxis on the road in 2020

Claim
Tesla would have robotaxis on the road in 2020, with coverage quoting more than a million.
Who made it
Elon Musk, at Tesla Autonomy Day.
Date
April 22, 2019.
Primary source
Fortune event coverage.
Independent check
Tesla’s Q3 2026 update reports Cybercab rides in Austin from September 3, 2026 and nearly 2 million unsupervised miles. TechCrunch found small fleets in a few Texas and Florida cities.
Still unproven
Fleet size and paid miles are not disclosed.
Status
Did not hold up
Last checked
2026-11-06

Read our check

AI would make radiologists obsolete

Claim
People should stop training radiologists because AI would beat them within five to 10 years.
Who made it
Geoffrey Hinton, at a Toronto machine learning event.
Date
2016.
Primary source
Eyewitness account by Joshua Gans.
Independent check
Fortune reports radiologist numbers up about 10% in a decade, thousands of open jobs and average pay of $571,000 in 2025. Hinton has since walked the claim back.
Still unproven
Whether AI tools change radiologists per scan over the next decade.
Status
Did not hold up
Last checked
2026-11-06

Read our check

GPT-4 scored in the top 10% on the bar exam

Claim
GPT-4 passed a simulated bar exam with a score around the top 10% of test takers.
Who made it
OpenAI, at the GPT-4 launch.
Date
March 14, 2023.
Primary source
GPT-4 Technical Report.
Independent check
A peer-reviewed re-evaluation puts GPT-4 near the 48th percentile among people who passed, and about the 15th on essays.
Still unproven
Exact percentiles depend on score data the exam body does not publish.
Status
Did not hold up
Last checked
2026-11-06

Read our check

Devin is the first AI software engineer

Claim
Devin can do end-to-end engineering work on its own, resolving 13.86% of SWE-bench issues unassisted.
Who made it
Cognition, at launch.
Date
March 12, 2024.
Primary source
Cognition launch post.
Independent check
Answer.AI found 3 successes in 20 tasks. Cognition’s CEO now describes Devin as a helper, not a replacement.
Still unproven
No independent test of how much real team work Devin finishes unaided. Revenue figures are company reported.
Status
Did not hold up
Last checked
2026-11-06

Read our check

The AI Pin would move people past screens

Claim
A clip-on AI device could replace much of what people do on phone screens.
Who made it
Humane.
Date
Shipped April 2024.
Primary source
TechCrunch.
Independent check
Engadget, citing The Verge, reported about 10,000 sold against a 100,000 goal. The Pin was shut off on February 28, 2025 after HP bought Humane’s assets.
Still unproven
Whether other AI wearables succeed where the Pin failed.
Status
Did not hold up
Last checked
2026-11-06

Read our check