![\"Colleagues](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-consciousness-debate-6-que-inline-1.jpg\")
<\/figure>\nWhat Rebecca Lowe’s Six Questions Are Actually Doing<\/h2>\n
Lowe writes at the ends don’t justify the means<\/em>, a Substack with over 2,000 subscribers covering “philosophical writing on freedom and all the other cool things.” She had planned something longer and more systematic. She scrapped it and shipped six questions instead, because the argument was already happening and a careful essay would have landed three weeks late.<\/p>\n That choice is the first lesson. Her stated goal is that discussion “can be both clarified and enriched by bearing in mind some simple distinctions that are ‘bread and butter’ to most philosophers.” Not resolved. Clarified. For anyone making procurement and process decisions, clarified is the only outcome that matters.<\/p>\n Lowe built the list so you can read it straight through or jump between questions based on the answers you give. A linear argument only works if everyone starts from the same premise. Nobody on your floor does.<\/p>\n Branching handles that. The engineer who thinks the model reasons and the inspector who thinks it pattern-matches take different routes and arrive at the same clarity about what each of them is actually claiming. This is how good acceptance criteria work too. You do not write one test and hope. You write a decision tree that sorts the system’s behaviour into categories you can act on.<\/p>\n Several of her questions exist to expose differences of view about what “the conscious thing” being referred to even is. Is it the model weights? The running inference? The chat persona? The company’s product wrapper around all three? People argue for an hour without noticing they are describing four different objects.<\/p>\n Run the same test on your own AI claims. When someone says “the system catches surface defects,” ask which system. The trained model, the camera rig, the threshold logic, or the operator who clicks accept. Most disputes about capability dissolve the moment you name the referent precisely.<\/p>\n Lowe is blunt about who dropped the ball. A small number of philosophers took jobs at AI companies, where “seemingly, they mostly spend their time supporting the views of their bosses.” The majority stayed in universities and, in her words, “spent these years telling each other how rubbish they think AI is, while complaining about how much harder it has become to mark essays.”<\/p>\n I strongly believe we’ve let the world down, as a profession.<\/p><\/blockquote>\n That matters to you because nobody shipped operating managers a usable vocabulary. You were handed two options: the vendor’s story or the sceptic’s shrug. So anthropomorphic language filled the vacuum, and it now sits inside your capital requests, your acceptance criteria, and the slides your supervisors use to explain a new system to second shift.<\/p>\n Watch the verbs in any AI inspection or predictive maintenance demo. The system “understands” the defect, “learns” your process, “knows” when a bearing is about to go. None of those words map to a testable specification. Try writing a factory acceptance test around “understands” and you will see the problem immediately: there is nothing to measure, so there is nothing to reject.<\/p>\n The replacement is boring and it works. Demand claims in the form of inputs, outputs, error rates, and failure conditions. What exactly does the model see? What does it output, in what units, with what confidence? Which conditions push it outside its validated range, and what happens then? A vendor who cannot answer that in plain terms is selling you a story about a mind rather than a tool with known limits.<\/p>\n The same discipline applies to internal messaging. If your change-management deck tells operators the system “thinks like your best inspector,” you have guaranteed both failure modes: blind deference when it agrees with them, and total dismissal the first time it misses something obvious. Describe the scope, the tolerance, and the override rule instead. Boring language produces calibrated trust.<\/p>\n Both failure modes start the same way: someone makes a claim about what is happening inside the model, then changes their verification behaviour based on that claim. Inner experience is unobservable. Output against spec is not. Every euro you lose here comes from confusing the two.<\/p>\n A team starts describing a vision-inspection model as if it “knows” what a defect looks like. Within a quarter, the manual sampling plan gets quietly thinned, because re-checking something that understands the part feels like bureaucratic theatre. Nobody writes this decision down. It just erodes.<\/p>\n Then the supplier changes a surface coating, or the line lighting degrades, and the false-negative rate climbs for six weeks before anyone notices. The model never understood anything. It matched patterns in a distribution that moved. The cost is not the drift itself, it is the detection lag you created when you retired the control that would have caught it.<\/p>\n The mirror image is cheaper to spot and more expensive to carry. A document-extraction agent gets labelled “just a parrot” in its first demo, usually by the most technically credible person in the room. The pilot dies. Three full-time people keep writing CAPA summaries by hand, and the decision is never revisited, because dismissal does not generate a review date the way a deployment does.<\/p>\n Run the arithmetic anyway. If the system clears your accuracy threshold on a held-out sample of real deviation records, the parrot question is irrelevant to the business case. Lowe’s point about the AI consciousness debate applies here exactly: the distinctions are “bread and butter” to philosophers, and skipping them leaves you arguing about metaphysics in a capacity meeting.<\/p>\n Treat trust calibration as a measurable variable. Define the spec, sample against it on a fixed cadence, and set the verification rate from the error rate you observe, not from how clever the thing sounds.<\/p>\nThe ‘choose your own adventure’ structure and why branching beats a linear argument<\/h3>\n
Naming the referent: what exactly is the thing you claim is conscious?<\/h3>\n
Why a Philosophy Gap Became an Operations Problem<\/h2>\n
How vendor framing exploits the vocabulary vacuum<\/h3>\n
![\"Empty](\"https:\/\/falcoxai.com\/main\/wp-content\/uploads\/2026\/10\/ai-consciousness-debate-6-que-inline-2.jpg\")
<\/figure>\nWhere Anthropomorphism Costs You Money on the Plant Floor<\/h2>\n
Over-trust: skipped verification and undetected drift<\/h3>\n
Under-trust: manual work nobody re-evaluated<\/h3>\n