Under AI Conditions 2.2.1-2

Fallibilism When the Information Environment Is Mediated

The self-correction practices in this unit — the assumptions register, the Weekly Review, the "I Changed My Mind Because…" log — were designed for a world in which the person doing the correcting has reasonable access to the evidence they need to evaluate. That world is changing.

AI-mediated information environments alter the conditions of fallibilism in three specific ways, each of which increases the cost of self-correction if left unaddressed.

Three Shifts

Provenance becomes opaque. When a claim arrives through an AI-generated summary, a recommendation algorithm, or a chatbot response, the evidence chain between source and claim is compressed or hidden. The weekly review's provenance step asks: What does this rest on? In an AI-mediated environment, the honest answer is often: I do not know, and the system is not designed to tell me. This does not mean the claim is wrong. It means the claim cannot be evaluated using the normal fallibilist method until provenance is restored — either by tracing the claim to its source or by finding independent verification.

The practice implication: when a significant belief rests on AI-generated or AI-curated information, the assumptions register should flag the provenance gap explicitly. "Source: AI summary — original evidence not independently verified" is not a disqualification. It is a calibration marker that tells you how much confidence the evidence can bear.

Update signals are filtered. Fallibilism depends on exposure to disconfirming evidence — the data that tells you your current position may need revision. AI recommendation systems are structurally optimized to surface content that matches your existing preferences and engagement patterns. This produces an environment in which confirming evidence arrives more frequently than disconfirming evidence, not because the world confirms your views, but because the filter is tuned to show you what you already believe.

The practice implication: the weekly review's update step should include a deliberate search for disconfirming information on at least one of the three reviewed items. The search does not need to be exhaustive. It needs to be intentional — a conscious act of seeking the evidence that the algorithm is least likely to surface.

Revision feels costlier than it is. AI systems that personalize content also personalize identity reinforcement. The more precisely your information environment reflects your stated views back to you, the more those views feel constitutive of who you are — which increases the psychological cost of revision. The assumptions register was designed to externalize beliefs so that revision does not feel like self-betrayal. Under AI conditions, that externalization is more necessary, not less, because the ambient information environment is actively working to internalize your beliefs as identity markers.

The Adjusted Practice

The core protocols — assumptions register, Weekly Review, revision log — remain sound. What changes under AI conditions is the intensity of three specific habits:

First, provenance tracking must become a default practice for any significant claim that arrives through an AI-mediated channel. The question is always the same: Can I trace this to a source I can evaluate independently?

Second, disconfirmation seeking must be a deliberate weekly review step, not an occasional exercise. The filter that shapes your information environment does not surface the evidence you most need to see.

Third, identity separation — the habit of holding beliefs on the page rather than in the personality — must be maintained with greater vigilance, because the environment is engineered to collapse that distance.

None of this requires rejecting AI tools. It requires practicing fallibilism with the awareness that the information environment is no longer neutral — and adjusting the self-correction protocol accordingly.