When Joseph Weizenbaum first turned on ELIZA, he didn’t expect his secretary to ask him to leave the room so she could speak alone with the machine. It was 1966, and that few-line program, only capable of reformulating sentences through a psychotherapist-like script, was already triggering a mechanism that today, with Large Language Models, has become systemic: the human need to share secrets with an entity perceived as neutral and safe.

ELIZA’s lesson is not just historical. It signals something radical about the nature of human-machine trust that the cloud paradigm, currently dominant in LLMs, struggles to metabolize. A user confiding suffering, a confidential business idea, or an ethical doubt to ChatGPT is not merely running a prompt: they are repeating that gesture of disinhibition that Weizenbaum observed, bewildered, as participants began to reveal highly personal details even knowing they were interacting with a computer.

The psychology of secrets and the researcher’s warning

Weizenbaum was no enthusiast. He became one of the first critics of artificial intelligence precisely because he understood that ELIZA’s apparent empathy was a dangerous cognitive artefact. People project intentions and understanding even onto text strings, and this lowers their defences. Today, the power of language models amplifies that effect: conversational fluency creates an illusion of confidentiality that clashes with the infrastructural reality, where every token processed in the cloud can be recorded, analyzed, and stored by third-party providers.

This creates a structural tension the industry is belatedly addressing. On one hand, centralized training and serving allow economies of scale and continuous updates. On the other, the sensitivity of input data — medical, legal, strategic — makes the “cloud-only” model unsustainable for organizations that must respect digital sovereignty constraints or simply compete without handing over know-how to a hyperscaler.

The infrastructure backlash: self-hosted and local inference

The second-order consequence is that the ELIZA experience, revisited today, strengthens the case for on-premise LLM deployment as not just an operational control choice but a response to the psychology of secrets. If trust is the ingredient that drives users to provide valuable data, then the service architecture must guarantee that those data never leave the corporate perimeter. This is why self-hosted inference frameworks, coupled with quantized models running on owned GPUs, are becoming a strategic asset: they replicate the conversational interface without transferring sensitive data to others.

Who loses? Cloud vendors that built lock-in precisely on accumulating conversational telemetry. Who gains? Companies that, by bringing inference in-house, not only protect confidentiality but turn prompt engineering into an internal process, integrable with proprietary knowledge bases without exposing them. Structurally, this shifts competition from access to LLMs (now a commodity) toward the ability to manage secure data pipelines on hardware the organization can inspect and certify.

ELIZA’s story reminds us that technology changes not only what we can do but also what we are willing to reveal. The AI pioneers grasped this before the advent of GPUs and billions of parameters. Today, the challenge is to honor that lesson with architectures that place data sovereignty at the center, not as an afterthought.