Originally published on LinkedIn, November 30, 2024.
We extend trust—like a line of credit—to machines, strangers, and corporations because we have come to expect reliability or reciprocity or responsibility in return. When we interact with AI agents, how can we avoid misplacing trust, creating risks we cannot yet measure or mitigate?
Unlike trust in machines, trust in AI agents isn’t based on deterministic performance. Unlike trust in humans, trust in AI agents isn’t based on commitment and goodwill. Unlike trust in corporations, trust in AI agents isn’t governed by contractual obligations. Not yet. Rather, our trust in AI agents is and ought to be based on their competence and their fidelity to our intentions and interests.
To make the stakes clearer, let me define two archetypes of personal AI agents—the factotum and the fiduciary—and draw a high-contrast distinction between the agents we have today and the agents we deserve.
What The Factotum?
The word factotum comes from the Latin facere (“to do”) and totum (“everything”). A factotum is a servant whose job is, quite literally, to do everything—a jack-of-all-trades, handling a variety of tasks that require competence but not specialization.
Factotums have had a rich history. In early modern Europe, households employed factotums for estate management, household errands, and just about anything that involved adaptable service. Over time, as industrialization prioritized specialization, the factotum’s star dimmed and they came to be seen as vainly overreaching their grasp. Malvolio in Shakespeare’s Twelfth Night exemplifies the archetype: eager and competent but ultimately self-serving and unreliable.
The personal AI agents of today are digital factotums—general-purpose assistants designed to perform low-risk, well-defined tasks. They excel at a broad range of remarkable tricks (just follow any AI influencer whose fragile mind was blown yet again by an announcement about the GenAI flavor of the month), but personal AI agents based on LLMs lack the depth to navigate high-stakes, complex decisions that require reasoning, empathy, and fidelity.
Factotum AI agents deliver undeniable utility but they are not trustworthy in the way human assistants might be. Their “loyalty” is to their developers, not to you. Can you really trust an investment AI agent or a healthcare AI agent or a legal AI agent or a banking AI agent the way you trust Alice, Bob, Charlie, and Jack with your personal interests?
This is where we, more like angels than the others, must fear to tread. Trusting a factotum AI agent beyond its competence and fidelity—allowing it to influence key decisions—creates unknown risks. Until we have metrics and mechanisms—like a trustworthiness score—we cannot trust AI agents for life-critical tasks.
The parallels between LLM apps and mobile apps are instructive. Of the 8.9 million mobile apps in the marketplace today, only 1% generate significant revenue, mostly in 5 categories—gaming, social networking, entertainment, shopping, and music.
The AI agents we have today excel in tasks that are just as… (I want to say trivial but I’ll bite my tongue) uncritical. Yesterday it was my co-pilot. Today it’s a cursor. Tomorrow it will be just some agent that I used to know.
However, the true potential of AI lies beyond these agents of mass distraction.
The mobile apps that help you do things today—in categories like business, medical, health and fitness, finance, utilities, education, navigation, and travel—are sadly in the long tail of revenue and downloads. But these are the precursors of an evolutionary leap; they are like the dormouse that survived the Chicxulub asteroid which drove the dinosaurs to extinction.
When the mobile apps in these categories are replaced by personal AI agents, which could do useful things on our behalf more competently, would we not be better served if we could trust them with our interests?
The factotum AI agent may be competent but is not trustworthy. What we need instead is a fiduciary AI agent that can act with care as well as competence, with fidelity to your long-term interests—a true partner in decision-making.
But until we can design and build fiduciary AI agents, we must learn to rely on factotum AI agents without over-trusting them.
Misplacing trust in a factotum AI agent is like trusting Igor from Young Frankenstein to help you perform a brain transplant. When you end up with the brain of Abby Normal, it’s hilarious in hindsight but catastrophic in consequence. These AI agents operate best when confined to low-risk tasks. Expecting them to handle complex, ambiguous situations is a recipe for betrayal.
In a Nutshell
Much has been written about the parallels between Frankenstein’s monster and AI, but I think the closer analogy for personal AI agents is Igor. The most dangerous AI agents are not those that fail spectacularly but those that succeed quietly without our comprehension, appropriating trust that they have neither earned nor deserve.
- Set Boundaries: Use factotum AI agents for routine, low-risk tasks where competence is the only requirement. Avoid relying on it for use cases involving trust. Even then, look for a quantitative measure of trustworthiness rather than a binary judgement.
- Question Alignment: Recognize that the incentives of your AI agents are set by its developers, not by you. Scrutinize its inputs, inner workings, and outputs accordingly.
- Demand Fiduciary AI Agents: As the field of AI continues to evolve, demand more from its developers. You deserve fiduciary AI agents—systems that act with both expertise and duty of care.
If your AI agent isn’t acting in your best interests, whose interests is it serving?
We’ll cut to the core of that conversation — vivisect Fiduciary AI agents –– in the next post.
