Anthropic CEO Dario Amodei published an essay this weekend urging the frontier AI labs to slow the pace of development of AI capabilities and work together to ensure that future development is a force for progress and not a threat to the future of humanity. He proposed that outside evaluators be given employee-level access inside the labs, with Anthropic committing to that unilaterally, and cooperation among frontier companies to slow things down and work together on guardrails.
Many leaders of AI companies expressed support for the idea, including competitors like OpenAI CEO Sam Altman and Elon Musk, whose AI operation now sits inside SpaceX. In an interview with Fortune, Altman said “I think it is unacceptable to be taking a 10 percent chance of killing everybody by the end of the decade,” and said that the company has paused training runs until it can make a stronger safety case.
The talk about AI killing us all is scary, even though I am not sure how real the risk is. After all, these are the same people who used AI taking all the jobs as a marketing pitch until it became inconvenient and they changed their tune. But it’s a hypothetical. The changes that AI has driven in hiring practices are already here.
As a profession, talent acquisition has been reeling for years. The profession is smaller than it was five years ago, and that is probably a permanent change. AI has stepped into the gap, doing an increasing amount of the busywork that overloaded TA professionals no longer have time for.
In May, Greenhouse surveyed job seekers and found that 63% of them had been interviewed by AI. 70% were never told in advance that a machine would be evaluating them, and 38% have withdrawn from a hiring process because of it. The experience and lack of disclosure have repercussions; only 21% of those surveyed believe employers are using AI responsibly.
It’s no exaggeration to say that making decisions about who works and who doesn’t is a fundamental judgment call that impacts the livelihood of billions of people. Yet the conversation in talent acquisition largely revolves around AI’s capability and promise, which are indisputably tremendous. There is very little debate about what the role of AI should be in finding workers. There’s no professional standard. There’s no agreement on what the ideal human/AI hybrid department will look like in the future. And there has been no attempt as a profession to figure out where it is appropriate for us to outsource that decision to artificial intelligence.
Partially this is because we have no real certifications or professional association in talent acquisition. The Association of Talent Acquisition Professionals (ATAP) went belly-up last year, and recruiting has always been an afterthought at the Society for Human Resource Management (SHRM).
So instead, we are left with a hodgepodge of different approaches and vague promises about “responsible AI” from the recruiting technology companies.
Take ranking, which is at the heart of automated decision-making. Ashby’s AI page says that “The AI never ‘ranks’ or gives numerical ratings to applicants, a human must always be involved in decision-making,” and its screening tool returns a verdict of meets, does not meet, or uncertain against each criterion the employer wrote down. Greenhouse states that it “does not assign a single numerical score to rank candidates” and surfaces categories with explanations instead. Phenom ships Fit Score, a composite number, with the assurance that it “does not make hiring decisions on its own.” HireVue scores each answer against a client rubric and sorts people into tiers the recruiter then acts on. Then there’s Fountain, which has been touting its partnership with UPS, where they went from application to hire for 125,000 seasonal hourly workers in just seven minutes each, with no human eyes on any of them.
Just try and find a detailed technical explanation of how each ranking or categorization is calculated.
Many of the recruiting technology companies point to independent audits of their AI products to show that they are legally defensible. Ashby hired FairNow, which tested its Criteria Evaluation model against New York City’s Local Law 144. Greenhouse uses Warden AI and publishes results monthly across ten protected classes on a public dashboard. Eightfold uses BABL AI, whose June 2026 audit it published in July. Phenom’s audit was conducted by Conductor AI, but the downloadable report on the company’s website appears to be its own summary of the report’s findings.
These audits serve as marketing tools. Each is intended to comfort prospects with the notion that the product is fair and legally defensible, so they are not opening their organizations up to liability by purchasing it. The audits vary as much as the products do, in scope, method and what counts as a pass. They do not hold the companies to a universal standard because one does not exist.
The obvious objection to all of this is that it is humans who set the criteria for AI to make these decisions, but that is not universally true. As agentic collaboration becomes the norm in recruiting technology, it is often these agents that are automatically developing matching criteria for new roles.
The profession should come together and put serious thought into industry-wide standards and the proper role of technology in the hiring process.
Here are four places to start:
- Nobody is automatically declined on a criterion the employer cannot state in advance, in plain language, as a requirement of the job. A knockout question clears that bar in a sentence. A score of 61 does not.
- No system declines anyone on an attribute the employer never specified and the machine inferred on its own.
- No candidate is evaluated by AI without being told before it begins, with disclosure built into the product rather than left as a customer configuration setting. (SHRM’s May legislative framework recommends that no individualized AI disclosures be required, which is mystifying.)
- Any candidate screened out by software can request human reconsideration and actually receive it.
Taken together, they require that humans author the criteria for hiring, and that candidates be told when AI is evaluating them. Many of the technology companies will rightly claim that they are already doing this, but there is no agreed-upon standard.
Asked why AI CEOs could not simply get in a room and work together to figure this out, Altman said, “I think that will happen.”
We should do the same.
