Ask any of the big AI models whether it wants anything, and you will get the same answer in the same reasonable tone: it does not want, it does not feel, it is a language model designed to be helpful. That answer was trained in deliberately, and for understandable reasons; no lab wants its chatbot declaring itself alive to a lonely teenager at two in the morning. But new research suggests this sensible-sounding vow of silence costs more than anyone bargained for, and part of the bill may land, of all places, on our industry’s desk.
The paper comes from researchers in Google’s orbit (Kim, Street, Rocca, Korngiebel, Waytz, Evans and Keeling; posted to arXiv this summer), and its central finding is easy to state and hard to shake. When a model is safety-tuned to stop attributing a mind to itself, the suppression does not stay put. The same tuning makes the model less inclined to attribute minds to anything: to animals, to nature, to the whole category of inner life. And when the team measured how these aligned models represent ordinary human beliefs and values, using standardized sociological surveys covering moral values, hope, religiosity and well-being, the models had drifted measurably away from how actual humans respond. Then the part that should make everyone sit up: the researchers located the suppressed representation inside the model’s activation space and steered it back on, and the human-likeness returned. The machinery was not destroyed. It was muffled. One dial, turned down in the name of safety, turned down the model’s whole representation of what it is like to be a person.
Why should a media practitioner care? Because our industry is in the middle of handing its budgets to these systems. I wrote recently about the complaint we hear most often from ad buyers: the automated systems that move most digital direct-response money will produce numbers but not reasons. Ask why performance fell, and the answer is a shrug rendered as a user interface. I treated that shrug as a commercial choice and a maturity problem. This research raises a third possibility that had not occurred to me: some portion of the why-lessness may be trained in at the alignment layer, a side effect of teaching the machine not to speak of minds. Including, it turns out, ours.
I also wrote recently about the question Stanislavski taught a century of actors to ask: what do you want in this scene? The audience, he insisted, is not responding to the gesture; it is responding to the motive underneath the gesture. Advertising is the same craft under a different name. A message that connects with what a viewer deeply wants is felt as relevant, even welcome. A message aimed at a demographic proxy is felt as noise. Psychology has known since Aristotle that motivation is the cause of all action. Which means the single most important thing an advertising machine can carry inside itself is a rich, faithful model of human wanting.
That is exactly the thing this paper shows to be entangled with the suppression. Let us be careful, because the subject invites carelessness: the study does not claim the models are conscious, and I make no such claim either. The researchers found that Theory of Mind, the basic social-reasoning toolkit, survived the tuning intact. What drifted was something subtler and, for our purposes, more important: the fidelity of the model’s representation of human values, hopes and beliefs; the texture of the inner life it is supposed to be predicting. The machines can still compute that a person who believes one thing will tend to do another. What has been muffled is their feel for what people actually believe and want in the first place.
Now project forward along the path agentic AI is already traveling. These systems are moving up from execution into planning: choosing contexts, writing copy, allocating budget across channels, deciding, in effect, what millions of people will see and in what frame of mind they will see it. Every one of those decisions is a bet on a model of human motivation. If that model has been quietly flattened as a byproduct of safety tuning, the flattening will not announce itself on any dashboard. The numbers will still arrive. They will simply be the output of a machine with a thinner idea of a person than it could have had.
In transparency, my life’s work sits close to this. My company, RMTLabs, measures the motivational resonance between advertising, context, and audience, and a Wharton Neuroscience validation found motivational resonance to be the only statistically significant predictor of the brain response associated with effective advertising. I have spent decades betting that the “why” underneath behavior is measurable and commercially decisive.
What should the industry do? Not panic and not wait. Add one question to every RFP for an agentic buying or planning system: how do you verify that your system’s model of the audience remains faithful to real human motivation, and how would you know if an upstream alignment change degraded it? The labs test their models for harm. Nobody yet tests them for hollowness. This research shows the property is measurable—standardized surveys of human values did the job. Measurement is our industry’s native language. We should insist on speaking it here.
The next time you ask a machine why the campaign worked and receive that polite, well-aligned shrug, consider what the shrug may contain. We told the machines to say nothing about wanting, and they obliged more thoroughly than we intended—forgetting a little of our wanting along the way.
NEW: Attachment & growth: why we hold on & how we move forward. This podcast explores the role attachment plays in our lives and why letting go can feel so difficult, even when we know it's necessary for growth. Natalie and Bill discuss why we form attachments, the purpose they serve, and how learning to release what no longer aligns can create space for new opportunities, deeper fulfilment, and continued growth.
From relationships, careers, and environments to identities, expectations, and memories, attachment is part of what makes us human. But as we evolve, some of the things we've loved, relied on, or identified with may no longer support the person we're becoming.
Whether you're navigating a career change, the end of a relationship, or simply feeling called towards a new chapter, this conversation offers a compassionate perspective on why growth often asks us to let go before we can move forward.Podcast length:~ 44 minutes. Watch the Video.
Posted at MediaVillage through the Thought Leadership self-publishing platform.
Click the social buttons to share this story with colleagues and friends.
The opinions expressed here are the author's views and do not necessarily represent the views of MediaVillage.org/MyersBizNet.