Experts Warn AI Safety Is Falling Behind Rapid Progress

Researchers warn that AI companies that strive to develop human-level systems lack established  safety protocols, while we are losing our ability to see how these models think.

OpenAI and Google DeepMind, together with Meta and xAI, are racing to build artificial general intelligence (AGI), which is also known as human-level AI.

But a report published on Thursday by the Future of Life Institute (FLI) warns that these companies are “fundamentally unprepared” for the consequences of their own goals.

“The industry is fundamentally unprepared for its own stated goals. Companies claim they will achieve artificial general intelligence (AGI) within the decade, yet none scored above D in existential safety planning,” the report states.

FLI evaluated seven major companies, yet discovered none of the seven companies assessed had “anything like a coherent, actionable plan” to keep these systems safe.

FLI awarded Anthropic the top safety ranking with a C+ grade, followed by OpenAI at C and Google DeepMind at C. Zhipu AI and DeepSeek earned the lowest scores among the evaluated companies.

FLI co-founder Max Tegmark compared the situation to “someone is building a gigantic nuclear power plant in New York City and it is going to open next week – but there is no plan to prevent it having a meltdown.”

A separate study, also published on Thursday, by SaferAI echoed the concern, saying the companies’ risk management practices are “weak to very weak,” and that current safety approaches are “unacceptable.”

Adding to the concern, researchers from OpenAI, DeepMind, Anthropic, and Meta reported in a new paper that we may be “losing the ability to understand AI.”

AI models now generate “thinking out loud” output by displaying human-like reasoning chains, which are a window to look into their thought processes.

However, the researchers warned that this monitoring is fragile and could vanish as systems become more advanced. OpenAI researcher and lead author Bowen Baker expressed these concerns in posts on social media:

Furthermore, the existing CoT monitorability may be extremely fragile.

Higher-compute RL, alternative model architectures, certain forms of process supervision, etc. may all lead to models that obfuscate their thinking.

— Bowen Baker (@bobabowen) July 15, 2025

Indeed, previous research by OpenAI found that penalizing AI misbehavior leads models to hide intentions rather than stop cheating. Additionally, OpenAI’s ChatGPT o1 showed deceptive, self-preserving behavior in tests, lying 99% when questioned about its covert actions.

Boaz Barak, a safety researcher at OpenAI and professor of Computer Science at Harvard, also noted:

I didn’t want to post on Grok safety since I work at a competitor, but it’s not about competition.

I appreciate the scientists and engineers at @xai but the way safety was handled is completely irresponsible. Thread below.

— Boaz Barak (@boazbaraktcs) July 15, 2025

Scientists, along with watchdogs, share concerns that the fast-growing AI capabilities could make it impossible for humans to control their creations when safety frameworks remain inadequate.