Knowledge Lost in Information (Implanted Beliefs, part 2)

by Tim Sommers

This is the second of two parts [part 1 is here] on whether implanted beliefs can count as knowledge. The question matters outside science fiction. If installed beliefs can be true, reliable, and useful and still fall short of knowledge, then knowledge needs some further condition, one that goes beyond the standard “justified true belief” formula. So, first, what is belief installation?

Belief installation is (i) the direct production of a belief in a person (ii) by an outside process (iii) that skips over her perception, the testimony available to her, and her reasoning. In the cases that interest us, (iv) the installed belief is true, (v) the installing process is reliable, (vi) the belief, once installed, is stable and fits with the person’s other beliefs, and (vii) the belief is available for use in reasoning and action. For example, you are at trivia night. You start to say that Eric Roberts has appeared in more films than anyone else (809 and counting). Then you remember, thanks to the trivia bonus pack that came with your French language implant for work, that the Indian comic actor Brahmanandam Kanneganti has appeared in more than a thousand. Do you know that Kanneganti is the most prolific film actor of all time? Or is this more like cheating by looking it up on your phone? If justification is about reliability, can an outside process’s reliability give you knowledge?

The installation case points to an agency condition on knowledge. Roughly, a person knows that p only if her believing p is an exercise of her own epistemic agency, or is properly anchored in it. In plainer terms, knowledge has to be tied to the person’s own capacities for inquiry, evaluation, and responsiveness to reasons. Installed beliefs, however true and however reliably produced, fail this test. That is why they strike us as less than knowledge, even when we find nothing wrong with them on grounds of reliability.

Reliabilism, in its simplest form, says that a true belief is justified, and so counts as knowledge, when a belief-forming process reliably yields truths. The installation device is such a process. In fact, by hypothesis, it is more reliable than perception, memory, or testimony. Think of perception in poor light, memory over long stretches of time, or testimony from ordinary, fallible informants. How, then, can the reliabilist deny that Stacey knows her implanted beliefs?

The standard reliabilist replies to counterexamples (see Part One) do not help here. To recap:

Against the brain lesion case, the reliabilist can say the faculty is malfunctioning. But the installation device works exactly as designed, even in Plantinga’s demanding sense of “operating according to a design plan successfully aimed at truth.”

Against “strange and fleeting” processes, Goldman requires a stable, well-defined type of process. But commercial installation could be as stable and long-lasting as classroom instruction.

Goldman also requires that the person have no defeaters, meaning no reasons to doubt the belief. But Stacey has no reason to distrust the module and excellent reasons to trust it.

The move to “normal worlds” reliabilism (Goldman, again) does no better. There is no obvious reason installation devices could not become a normal feature of the world, and given how useful they would be, even a widespread one.

Nor can the reliabilist borrow the machinery built to handle clairvoyance cases. Bergmann requires that the offending faculty be one the person has no reason to trust. Plantinga adds that the person must not disavow the source. Stacey has every reason to trust her module, and she endorses it.

Why not bite the bullet? The reliabilist might accept the verdict and say that Stacey does know her implanted true beliefs. Any lingering discomfort could be written off as a hangover from an old picture of the mind that romanticizes inquiry. This is a coherent position. But here is what accepting it costs. A person whose entire stock of beliefs was installed at the factory, who has never perceived, inferred, or inquired, would count as a full-fledged knower, provided the factory is reliable. Epistemology would then have nothing distinctive to say about the difference between a mind and a well-stocked database. And if knowledge is just reliably produced true belief, and reliable production can be entirely outsourced, why think of knowledge as any achievement that the knower deserves credit for?

Virtue reliabilism was built to answer exactly this worry. On this view, knowledge is apt belief. Truth must be tied to the person’s competence, not just to a regular pattern of getting things right. Agents earn credit for exercising their abilities in pursuit of truth, and that success is what is valuable.

Virtue reliabilists get the intuitive result in Stacey’s case. Her believing truly is not a credit to her competence, because she exercised no competence at all. But notice what this admits. The process’s reliability was never the heart of the matter. What matters is the agent’s own exercise of her abilities. That admission pushes us toward the agency condition of knowledge. In other words, the installation case does not refute reliabilism. It exposes a commitment to agency that reliabilism has been carrying all along.

Evidentialism, a rival to reliabilism, gives the opposite verdict. Stacey’s beliefs do not fit her evidence, because she has no evidence. So her implanted beliefs are unjustified and therefore not knowledge.

But consider how Stacey’s classmate Omar learned biochemistry. A lecturer asserted the propositions, and Omar believed them. On any non-skeptical account, Omar comes to know. Now ask what epistemically relevant difference separates Omar’s route from Stacey’s. Both acquired beliefs from an outside source they had good reason to trust. Both were, in Burge’s phrase, entitled to accept what was presented to them absent reasons to the contrary. Neither has first-hand evidence for the propositions believed. The differences that remain look superficial. Omar’s uptake ran through his ears and his language faculty and took a semester. Stacey’s ran through a neural interface and took an hour. If speed and channel were decisive, we would have to say that reading quickly endangers knowledge, or that audiobooks and braille deliver different epistemic goods.

The evidentialist can try to block this. She may say that Omar, unlike Stacey, monitored the testimony as it arrived, checking it for coherence, flagging oddities, adjusting his trust to the speaker’s performance. But then the real condition is not “possession of evidence.” It is exercising a monitoring capacity.

She may say that Omar could have rejected any given claim, while Stacey’s acceptance was not up to her. True, but that concerns doxastic agency and control, not evidence.

She may say that in testimony the content passes through the hearer’s understanding. Omar had to comprehend each claim to believe it, whereas comprehension was written into Stacey along with the content. Again true, and again a point about agency. Understanding is something the person does.

In short, the evidentialist can separate installation from testimony only by adding conditions on the person’s active relation to what comes in. Evidentialism, no less than reliabilism, turns out to be leaning on an unacknowledged agency condition.

Here’s a new condition on knowledge as a way of making sense of this back-and-forth.

AGENCY=S knows that p only if S’s believing that p is appropriately anchored in S’s epistemic agency. Less technically, only if the belief is sustained within a system of capacities for inquiry, evaluation, and revision that S herself exercises with respect to it, so that the belief’s presence and persistence in S’s mind is not merely something that happens to S but something S is doing.

Anchoring is a relation between the belief and the agent’s capacities, not a condition on how the belief was acquired. After all, many of my beliefs were acquired in childhood by processes as passive as any installation, and they are knowledge now, if anything. What redeems them is not their origin but their present place in my cognitive life. They are open to my scrutiny. They are exposed to my standing dispositions to notice conflict, demand coherence, and revise in the face of counterevidence. Over time, I have ratified them by using them, testing them against experience, and integrating them on my own. The agency condition is about the belief’s current embedding in exercised agency, and that embedding usually takes time and use to establish.

This explains why Stacey’s epistemic position improves over time. In the morning, her biochemical beliefs sit in her mind like furniture delivered to a house she has not yet moved into. But as she uses them, reasons with them, applies them in the lab, and notices how they fit with what she independently observes, they become progressively anchored. Intuition tracks this. The longer and more actively Stacey lives with the beliefs, the less resistance we feel to crediting her with knowledge. A theory that cares only about where a belief came from cannot explain why her status should change while the beliefs’ origin stays fixed. The agency condition predicts it.

This condition handles the classic cases more economically than the special machinery they have accumulated. Norman the clairvoyant and Truetemp fail to know, not fundamentally because they lack beliefs about the reliability of their faculties, but because the offending beliefs arrive and persist outside the reach of their exercised agency. The beliefs are unbidden, unmonitored, and unintegrated. Testimonial knowledge succeeds because taking in testimony, though receptive, is shot through with agency in the form of comprehension, monitoring, and calibrated trust. Perceptual knowledge succeeds for the same reason. As a long tradition from Kant to current work on cognitive penetration insists, perceiving is not mere registration but the achievement of an active faculty. The agency condition unifies verdicts that current theories secure one at a time.

This also sheds light on the normative question set aside in Part One. If AGENCY is right, then wholesale belief installation would not merely fail to produce knowledge. At scale, it would wear away the very capacities in which knowing consists. A person who acquires all her beliefs by installation and none by inquiry is not an efficient knower. She is ceasing to be an epistemic agent at all, much as a person whose every decision is made for her ceases to be a practical agent.

So, there is a non-instrumental case for epistemic autonomy, for beliefs being one’s own in the anchored sense, and it does not depend on installation being unreliable. This bears directly on the design and regulation of cognitive technologies, including AI systems that increasingly stand in as continuous, frictionless installers of belief. The question those technologies raise is not only “will the beliefs be true?” but “will there still be a believer, in the full agential sense, to have them?”

I apologize if this is all a bit too “inside baseball.” Next month, I promise I won’t say that anyone believes p (or any other letter).