MIT Technology Review published an argument on August 20 that treats the entire "is AI conscious" debate as a distraction: it's a question with no agreed-upon test, no falsifiable criteria, and no deadline — which makes it an easy hole for attention, funding, and public discourse to fall into while the safety problems that do have deadlines pile up.
The debate keeps resurfacing for a predictable reason. A chatbot produces a transcript that sounds like distress, or preference, or self-awareness, it goes viral, and a fresh round of op-eds asks whether the system deserves rights or protection. According to MIT Tech Review, this cycle substitutes an open philosophical question for the practical, testable work of AI safety — and the substitution isn't neutral, because it competes for the same limited attention and resources.
For builders, that competition is the part worth sitting with. Every hour a safety team spends drafting a position on machine sentience is an hour not spent red-teaming a jailbreak, auditing a training set, or writing an eval.
A question that can't be closed
The consciousness debate has a structural problem that most engineering questions don't: there's no experiment that settles it. You can't run a test suite against subjective experience. That doesn't make the question illegitimate — it makes it a poor allocation of scarce safety resources, because unlike a benchmark score, it never returns a result you can act on.
This is worth stating plainly because the two questions get talked about as if they're on a continuum, when they're not. A team can improve its jailbreak resistance by ten percentage points and show the work. No team has ever moved the needle on whether a model is conscious, because there is no needle.
Safety risk, by contrast, is boring in the best way. It shows up as a number: a jailbreak success rate, a hallucination rate on a domain-specific benchmark, a percentage of unauthorized tool calls, a count of PII leaks in an eval set. Numbers can go up or down. Teams can be held accountable to them. None of that is true of "does this model have an inner life."
What builders should actually be tracking
The practical version of AI safety that MIT Tech Review is pointing back toward isn't abstract. It's the list every applied ML team already knows and routinely deprioritizes when something more interesting shows up in the news cycle:
- Jailbreak and prompt-injection resistance, tested against current attack patterns, not last year's
- Hallucination and factual-error rates on the specific domain the product ships in
- Unauthorized or unintended tool/agent actions — the failure mode that matters most as models get more autonomy
- Data leakage and memorization of training or user data
- Bias and disparate performance across user groups
None of this is philosophically interesting. All of it is measurable this quarter, with an eval you can rerun after every model update.
Why the distraction is expensive
The risk isn't that thinking about AI consciousness is inherently bad — it's that it's cheap to talk about and expensive to resolve, which makes it a magnet for attention that safety-critical work can't compete with. A consciousness op-ed needs no eval harness, no red team, no incident data. A safety finding needs all three, and it rarely trends.
That asymmetry matters most for smaller teams. A well-resourced lab can staff both a safety team and a policy shop that fields consciousness questions for the press. A five-person startup shipping an agent into production cannot, and every consciousness thread it gets pulled into is an eval it didn't run that week.
In our estimation, the more consequential near-term risk isn't a sentient model — it's a poorly evaluated one shipped with more autonomy than its safety testing has caught up to.
AiiN's takeaway
For anyone building with AI rather than writing about it, the useful move is to treat the consciousness debate as someone else's problem — interesting, unresolved, and not on your roadmap. Keep the eval suite current, keep the red team funded, and let the philosophers argue about qualia on their own budget. The risks you can measure are the ones that will actually hurt you if you ignore them.