If the latest AI Safety Index isn’t the bucket of cold water you needed about the people shoehorning GenAI into every aspect our lives while blowing a multi-trillion dollar hole in the global economy, I don’t know how to help you.

Even the most generous reading of these scores reveals an industry that has lost even a shred of moral hazard and frankly, looks like another example of privitising profits while gearing up for a public bail out. The best company gets a C+. The rest of the scores of Cs, Ds, and Fs read like a school report for a child who spent the term eating glue.

Three companies just failed outright. And this is not for “todo” apps or support bot. The companies that failed are building frontier AI systems being shoved into work, education, medicine, defence, public services, and the daily cognitive bloodstream of society. And if you thought that at least they were taking their role in society seriously, think again. From the report:

“Even the three strongest performers – Anthropic, OpenAI, and Google DeepMind – were found to have “no explicit safety strategy,” scoring above the rest only because they have had more substantial track records of research in the 214 results related fields and support for external researchers.” (Emphasis mine…)

For testers, my message is brutal: stop carrying water for these companies and the AI Test Bros and do your goddamn job!

Evaluating AI like it is a magic search box is simply just not good enough. These companies have weak safety frameworks, poor independent audit, and a widening gap between safety rhetoric and actual behaviour. We are meant to be poking holes in unverifiable vendor claims not admiring the technology, and they aren’t even trying to pretend that they have a functioning risk management process:

“Yet even for the strongest players, reviewers identified structural gaps that begin with how risk itself is defined. One panellist found an “overly narrow definition of dangerous capabilities, accompanied with no discussion of how risks are chosen,” and warned that the prevailing focus is “out of touch with the reality of risks (including death).” This narrowness invites a “safety-washing” critique – companies “pushing their product as safer than alternatives” when, for the largest risks, the differences are marginal.”

I know we play the hand we’re dealt in life, but having been in the game for a while, trust me, how you play those cards right now in testing will be a defining moment in your career.

Enabling AI in your system should immediately place the risk as HIGH, and if you are involved in testing AI, setting benchmarks for safety is the FLOOR and only proves it somehow survived a narrow inspection under artificial conditions.

The information you’re testing for should include “what can this model help someone do faster, cheaper, and at scale?” like fraud, manipulation, cyber abuse, self-harm reinforcement, bad medical advice, and industrial-strength stupidity. That’s because the tools, workflow integrations, internal deployment, and basic human desire to trust these systems are not the context in which they’ll be used.

Operational controls also need to included in your systems tests now with HARD thresholds. What blocks release? What triggers rollback? Who can stop deployment? If nobody can answer, you do not have a functioning safety framework. As well, the entire organisation should be in scope in some way including whistleblowing, incident response, auditability, transparency, governance, and commercial incentives as part of the system.

Think I’m over-reacting? The conclusion of the report should be sobering for anyone claiming people pushing back against the avalanche of AI BS being served as serious work as “AI sceptics”.

“Against these rising stakes, the review panel found that companies have not kept pace; and in several respects, they argued, companies are moving backward. According to the panellists, no industry leader has a credible plan to control the increasingly autonomous and capable systems it is building. The thresholds that are meant to govern critical development and deployment decisions are kept qualitative rather than holding specific and firm. And public safety rhetoric increasingly diverges from commercial and political conduct.”

“But what reviewers considered the most troubling of all is that the leaders racing hardest toward the capability frontier – Anthropic, OpenAI, Google DeepMind, and Meta – have weakened or voided their prior commitments to pause, including replacing unilateral pledges with competitor-contingent conditions. In the panel’s assessment, this dynamic has undermined safety frameworks across the board and incentivizes a collective race to the bottom, rather than to the top.”

Race to the bottom indeed…