The AI Effect (Defining AI)

Context: FIT1061_MOC · why the unit drops the label and classifies by algorithm instead — the successor move to Turing Test (Imitation Game) What it weighs: whether “AI” picks out any stable property of a system, or only its novelty at the moment of asking.

Quick Revision

• 🎯 Test question: does calling this system “AI” change anything you can predict about it?
• ⚠️ Key Constraint: the boundary is not a property of the system ➔ answer “is it AI?” by naming the machinery and the failure mode, never by defending a label.

📝 Core Commitments

• McCarthy’s definition ➔ 1956, coining the term: “AI is the science of making machines do things that would require intelligence if done by humans.” ➔ keyed to task difficulty for humans, not to any machinery.
• The AI effect ➔ McCarthy again: “As soon as it works, no one calls it AI anymore.” ➔ the label tracks unfamiliarity, so every success removes a system from the category.
• Restated 70 years on ➔ Narayanan & Kapoor, AI Snake Oil (2024): “AI is whatever hasn’t been done yet.”
• The boundary sits nowhere ➔ spellcheck was AI in the 1980s and is not now while doing the identical thing; Roomba plans but never learns; AlphaFold solved a 50-year-old open problem yet reads as “not AI” to most people because it does not converse.
• What the label refuses to track ➔ novelty, conversational surface and marketing move it; capability, mechanism and risk do not.
• The unit’s resolution ➔ classify by the question the algorithm answers ➔ AI Algorithm Blocks (Search, Uncertainty, Learning); “intelligent” and “AI” are used as history, never as definitions.

⚖️ AI Boundary Spectrum

(The ten W1 systems. The columns separate cleanly; the label does not.)

System	Machinery as taught	Learns from data?	Plans / searches?	Unit block
Mechanical thermostat	bimetallic coil tilts a mercury switch — no code at all	✗	✗	—
Nest learning thermostat	updates internal parameters from your behaviour	✓	✗	C (W10–11)
Google Maps routing	shortest path on a graph; algorithm published 1959	✗	✓	A (W2–5)
Word spellcheck	dictionary lookup + edit distance to nearest match	✗	✗	— (was AI in the 1980s)
Spam filter	estimates $P(\text{spam}\mid \text{words})$ from past examples	✓	✗	B (W7)
Roomba	maps the room, plans a path — never learns	✗	✓	A
Tesla Autopilot	camera + deep network steers	✓	✗	C (W11)
AlphaFold	predicts 3D protein structure; 2024 Nobel	✓	✗	C (W10)
COMPAS recidivism score	$137$ inputs → risk score, used in US courts	✓	✗	B (W9)
LLM writing the lecture notes	next-token prediction, gradient-trained, agent-wrapped	✓	✓	A + B + C

When It Flips: nothing in the two capability columns moves when the label does ➔ the label is tracking the calendar, not the machine.

🧩 Case Application Drill

Case 1: your bank markets its fraud detector as "AI-powered". A colleague insists it is "just statistics". Settle it.

Model verdict

• Claim on trial: that “AI” vs “just statistics” is a factual disagreement about the system.
• Framework applied: under McCarthy the task qualifies (a human analyst spotting fraud would need judgement); under the AI effect it stops qualifying the moment it ships and works. Both verdicts follow from the label’s rules, not the detector’s.
• Verdict: the disagreement is unresolvable and unimportant. The answerable questions are: it estimates $P(\text{fraud}\mid \text{features})$ from past labelled cases (Block B), it fails on fraud patterns absent from its training data, and it affects the customers it freezes out. Marketing pressure pushes the label up; the AI effect pushes it back down.
• Counter-position answered: “the label matters for regulation” ➔ then the regulation must name the machinery and the affected population, exactly because the label is a moving target.

Case 2: spellcheck was uncontroversially AI in the 1980s and is uncontroversially not AI today. Its algorithm never changed. What changed?

Model verdict

• Claim on trial: that the category “AI” tracks a property of systems.
• Framework applied: McCarthy’s own observation — the category is defined relative to what machines are not yet known to do.
• Verdict: the expectations changed, not the system. Dictionary lookup plus edit distance was a machine doing something only careful humans did; once ubiquitous, it became “how spellcheck works”. This is the AI effect in its cleanest form and it is why the unit indexes by algorithm.
• Counter-position answered: “so today’s LLMs will stop being AI too” ➔ on this evidence, yes — which is precisely the argument for describing the machinery instead.

⚠️ Common Mistakes

• 💡 Defending the label ➔ arguing whether a system “really is AI” answers nothing; state the machinery, the failure mode, and who it affects.
• 💡 Conversational bias ➔ treating chat interfaces as the criterion, which misclassifies AlphaFold and Google Maps while flattering any chatbot.
• 💡 McCarthy quoted flat ➔ citing his 1956 definition without his own later retraction misses that he named the instability himself.