Topics
Latest
AI
Amazon
Image Credits:Tomohiro Ohsumi / Getty Images
Apps
Biotech & Health
Climate
Image Credits:Tomohiro Ohsumi / Getty Images
Cloud Computing
Commerce
Crypto
endeavour
EVs
Fintech
fund-raise
widget
Gaming
Government & Policy
Hardware
Layoffs
Media & Entertainment
Meta
Microsoft
seclusion
Robotics
security measures
Social
distance
startup
TikTok
Transportation
speculation
More from TechCrunch
Events
Startup Battlefield
StrictlyVC
Podcasts
Videos
Partner Content
TechCrunch Brand Studio
Crunchboard
Contact Us
On Monday , OpenAIlaunched a young mob of AI model , GPT-4.1 , which the caller said outperformed some of its live models on certain tests , particularly benchmarks for computer programing . However , GPT-4.1 did n’t ship with the safety report that typically accompanies OpenAI ’s manikin releases , known as a model or system card .
As of Tuesday morning , OpenAI had yet to issue a safety report for GPT-4.1 — and it seems it does n’t be after to . In a statement to TechCrunch , OpenAI voice Shaokyi Amdo said that “ GPT-4.1 is not a frontier fashion model , so there wo n’t be a separate organization card released for it . ”
It ’s fairly standard for AI science lab to release safety reports showing the types of tests they take internally and with third - company partners to evaluate the safety of peculiar models . These reports occasionally reveal unflattering information , like thata mannequin lean to cozen humansor isdangerously persuasive . By and large , the AI community perceive these report as undecomposed - religious belief attempt by AI labs to brook main research and red teaming .
But over the past several months , leading AI labs appear to have lowered their reporting standards , prompting rebound from safety research worker . Some , likeGoogle , havedragged their feeton rubber reports , while others have put out reportslacking in the usual detail .
OpenAI ’s recent track record is n’t special either . In December , the fellowship draw off literary criticism for releasing a safety reportcontaining benchmark results for a manakin differentfrom the version it deployed into production . Last calendar month , OpenAIlaunched a model , thick research , weeks prior to publishing the system board for that model .
Steven Adler , a former OpenAI safety researcher , note to TechCrunch that safety reputation are n’t mandated by any law or regularization — they ’re voluntary . Yet OpenAI has made several commitments to government to increase transparence around its poser . forrader of the U.K. AI Safety Summit in 2023 , OpenAI in a blog postcalled organisation cards“a key part ” of its approaching to accountability . And conduct up to the Paris AI Action Summit in 2025 , OpenAI said scheme cards leave worthful insightsinto a model ’s risks .
“ System cards are the AI industry ’s independent creature for transparency and for describing what rubber testing was done , ” Adler told TechCrunch in an electronic mail . “ Today ’s transparentness norm and commitments are finally voluntary , so it is up to each AI company to resolve whether or when to release a arrangement card for a pass modeling . ”
Join us at TechCrunch Sessions: AI
Exhibit at TechCrunch Sessions: AI
GPT-4.1 is shipping without a system card at a prison term when current and former employee are raising worry over OpenAI ’s safety exercise . Last week , Adler and 11 other ex - OpenAI employees file a proposed amicus brief in Elon Musk ’s case against OpenAI , indicate that a for - profit OpenAI might cut corners on safety work . The Financial Times lately reportedthat the ChatGPT Jehovah , spurred by militant pressure , has slash the amount of time and resourcesit allocates to safety quizzer .
While GPT-4.1 is n’t the highest - performing AI poser in OpenAI ’s roll , it does make real gain in the efficiency and latency departments . Thomas Woodside , co - founder and policy psychoanalyst at Secure AI Project , assure TechCrunch that the operation advance make a refuge account all the more decisive . The more advanced the example , the gamey the risk it could pose , he suppose .
Many AI labs have batted down efforts to codify safety equipment coverage requirements into jurisprudence . For example , OpenAI oppose California ’s SB 1047 , which would have take many AI developers to audit and publish safety rating on model that they make public .
