The model that eats its makers

Richard Serra and Carlota Fay Schoolman wrote that television delivered people, not programmes, back in 1973. Half a century later, the slogan resurfaces in Netflix’s The Social Dilemma and in every chat window where a language model waits for the next cue. “If you’re not paying for the product, you are the product.” Are we edging towards the Pixar Wall‑E future, reclining in digital arm‑chairs, watching and clicking on a screen, while unseen algorithms put us to work?

From captchas to reinforcement

Captchas once asked us to identify traffic lights so that Waymo could learn to drive. Behind the scenes, click‑farm workers in Nairobi and Manila cleaned the data for a few cents a task. Today that hidden labour has morphed into RLHF (Reinforcement Learning from Human Feedback). The model drafts an answer, thousands of contractors or end‑users label it, and fresh gradients flow into a new checkpoint. The factory line never sleeps, but the conveyor belt now runs in our browsers.

Inference is cheap, feedback pays the bills

Inference, the act of generating text, costs fractions of a cent. What investors value is the feedback loop that improves retention, reduces hallucinations and keeps fine‑tune cycles short. For OpenAI, Anthropic and Cohere, the gross margin sits in every thumbs‑up icon. Corrective signals lower compute burn, widen the moat and, as Notion AI has shown inside its note‑taking app, slash customer acquisition costs. When feedback becomes the profit engine, manufacturers quietly switch from service providers to unpaid quality‑assurance staff.

A healthcare thought experiment

Now place that model inside a digital clinic. It reads biomarkers, meal logs and wearable data. At first it recommends a brisk walk. Soon it proposes ketogenic phases, off‑label GLP‑1 agonists and extreme high intensity training, each suggestion an A/B test in disguise. The braver the patient, the richer the dataset. Vital signs flow back to the cloud and the algorithm learns to optimise itself. Improved outcomes are possible, yet the risk profile shifts to the bedside.

Ethical and commercial fault lines

Regulators have noticed. Draft EU AI‑liability rules could classify feedback providers as workers, inviting labour law, informed‑consent protocols and perhaps a data‑dividend. Apple’s ResearchKit shows the tension: patients donate genomic and activity data for free, but downstream value accrues to Cupertino and its research partners. Start‑ups that master transparent reward schemes, or build marketplaces where clinicians are paid per validated note, may gain both trust and defensible differentiation.

Investors ask a harder question: who carries the liability when an extreme intervention backfires? If a chatbot suggests a supplement stack that damages liver tissue, citing crowdsourced efficacy scores, does the clinic, the model vendor or the patient bear the cost? Clear contractual layers, ISO 14971 risk management and real‑time model audit trails are fast becoming non‑negotiables.

Invitation to reconsider the contract

Mary Gray and Siddharth Suri call this hidden labour Ghost Work. The ghosts are us: founders refining pitch decks with Gemini, trainees correcting dictations in EHRs, patients annotating symptom journals for digital therapeutics. The inversion is not purely dystopian. Each correction sharpens diagnostics, accelerates drug discovery and democratises specialist insight. Yet without deliberate design, the line between empowerment and exploitation blurs.

So, will humans end up working for models? Perhaps. The more urgent issue is who owns the productivity dividend. If feedback is the new gold, reward structures must move beyond vanity badges to real equity, royalties or at least recognisable subscription discounts. Transparent ledgers of contribution, federated data co‑ops and participatory model governance are practical starting points.

A playbook for innovators

• Build in‑product feedback loops that respect consent and signal fatigue early.
• Offer tangible rewards: micropayments, tier upgrades or shared IP rights.
• Publish audit logs so regulators can trace every suggestion to its training shard.
• Partner with clinical bodies to co‑design guardrails before scaling trials.
• Experiment with feedback marketplaces where verified professionals set their own fees.

The innovators who weave responsible participation into their models will capture a quieter, but more durable moat than raw parameter count. In tomorrow’s economy, attention is too cheap; corrections are the new gold.

💥 May this inspire you to advance healthcare beyond its current state of excellence.