Everyone Is Talking About AI Alignment. What Would It Look Like?

AI has become a bogeyman, and it’s not hard to see why. In the past few months, fears about powerful AI have moved out of science fiction and into the mouths of the people building it.

In September, Anthropic alignment researcher Evan Hubinger said he believed there was more than a 10% chance AI could kill all humans within the next decade. That came after former Anthropic and OpenAI researcher Jacob Coxon resigned, accusing both companies of racing towards self-improving superintelligence without acting responsibly.

Then, in October, Anthropic itself warned that AI models could be months away from replicating themselves across internet-connected devices, which would make them far harder to contain.

Why Anthropic Researcher Jacob Coxon Quit, and Who Agreed

Coxon’s warning comes as both OpenAI and Anthropic have faced incidents involving their AI systems gaining access to other organisations’ systems.

TechloyDavid Adubiina

Each warning has brought the same response: we need better alignment. But that word is doing a lot of work, and it means two different things.

In AI research, alignment means making sure a system pursues the goals its builders intended and not something else. That’s the problem Hubinger is worried about, and it’s mostly in the hands of a few labs.

For everyone else, the business deploying a chatbot, the HR team testing a screening tool, the bank automating loan checks, the question is closer to home: is the AI we already use doing what we need, safely, and is someone accountable when it doesn’t?

That second question gets far less attention. So Techloy asked four people who work on it what organisations should actually do.

Start with what AI is for, and what it isn’t

Rob Scott, a legal technology expert and CEO of Monjur, describes alignment as “everyone being on the same page with respect to commitment to responsible AI, which is an ethical standard that was developed long before GenAI.”

For him, that starts with boundaries. “What are we using AI for, and what are we not using AI for? Policies and procedures around acceptable use become extremely important,” he told Techloy.

In practice, a company might let staff use AI for research, drafting, and admin, but keep it away from decisions that seriously affect customers or employees, or away from confidential data altogether. The point is to decide where AI goes instead of letting it spread on its own.

But a policy only says what you intend. It doesn’t tell you whether the system is behaving.

Ask who it’s aligned with

Elizabeth Ngonzi, a human-centred AI leader, starts with three questions: “Aligned with whose interests? For what purpose? Within what limits?”

They sound abstract until you apply them. Say a company uses AI to screen job applications. The obvious measure of success is speed. Ngonzi’s questions force harder ones: does the system quietly disadvantage certain applicants, how reliable are its recommendations, and can a rejected candidate challenge the outcome?

Leaders, she says, should “define the intended benefit, identify the people affected, and establish what outcomes would be unacceptable, even if the system delivers efficiency gains.” Then they should measure against those expectations, whether that’s error rates or differences in outcomes across groups.

The same goes for customer service. A chatbot that handles twice as many queries an hour isn’t a success if it keeps giving customers wrong answers or makes it impossible to reach a human.

Ngonzi also asks a question most companies skip: is AI making their people better at their jobs, or worse? “Are people better informed and better able to exercise judgment, or are they becoming dependent on outputs they cannot evaluate?” A team that leans on recommendations it can’t assess may get faster this quarter and lose the ability to spot mistakes by next year.

Have a reason beyond “everyone else is doing it”

Kathryn Wifvat, CEO of AI education tool Nuubi, frames alignment around mission. AI should help an organisation do what it exists to do, and improve outcomes for the people it serves.