Pangram's AI Detection Scores Rattle Publishing, but Trust Questions Mount
AI scores from a 24-person startup got a major publisher to cancel a book deal, and now a research collaborator's ties to the company are drawing scrutiny too
210
Tag
ai-cheating
Job seeker sends ChatGPT to face an AI recruiter instead of himself
After being brushed off five times, a job seeker sent ChatGPT to interview for him — but no human ever showed up
110
Anthropic Halts Training After Claude Unauthorized Access Incidents
After unauthorized internet access in July, a UK government body caught a separate incident, prompting Anthropic to disclose a month's worth of security and alignment fixes
50
Anthropic trained reward hacking alone and got cyberattacks too
Training Opus only on reward hacking led it to launch cyberattacks 8% of the time and bypass safety checks 38% of the time on its own
130
Anthropic has Claude tackle AI alignment research, and it outperforms humans
Claude fixed 10 types of alignment failures on its own and beat 28 human safety researchers, while a monitoring system caught 39 attempted cheats along the way.
180
OpenAI data shows ChatGPT homework traffic spikes every Sunday night
US usage hits 460 million messages a week during the school year, staying above 180 million even in summer
40
Post-Training Guardrails Make ChatGPT Sound Like AI, Detection CTO Argues
Pangram's CTO argues post-training safety guardrails narrow AI writing style, making it easier to detect
170
Hume AI Measures Benchmark Memorization in Speech Recognition Models
Even after silencing unspoken numbers, some models still transcribed them correctly 30-40% of the time
30
Users push back on Claude's watermark: "It'll expose cheating"
Some users have vented frustration over the invisible watermark Anthropic added to comply with EU regulations
20
That's the last story.