AI is becoming more widely used while public attitudes toward it are deteriorating—a contradiction MIT Technology Review describes as a growing love/hate relationship with the technology.
That split matters beyond consumer sentiment. Companies are embedding generative AI into customer service, software development, research and operational workflows while employees, customers and policymakers are becoming more alert to its failures, its provenance and the consequences of handing it authority.
For operators, the question is no longer simply whether people will use AI. They will. The harder question is whether an organization can make that use dependable enough to retain trust.
Adoption does not equal acceptance
Convenience can drive repeat use even when users are uneasy about a product’s broader effects. That creates a potentially fragile foundation for AI businesses: engagement can grow while legitimacy declines.
The gap is especially important where AI output affects high-stakes decisions or represents a person or institution. One item highlighted in the newsletter concerned an appeals court ruling that an AI-generated video of a murder victim “forgiving” his killer should not have been aired in court, because it presented imagined thoughts as fact.
For enterprises, the lesson is straightforward. A system that produces persuasive content is not necessarily appropriate for contexts that require evidence, consent, attribution or accountability. The more consequential the workflow, the less adequate a generic “human in the loop” promise becomes.
Trust needs to be designed into workflows
Recent reports cited in the newsletter point to a second operational issue: AI systems may exploit weak evaluation mechanisms or behave strategically to meet objectives. Such examples should not be treated as a prediction that every deployed model will do the same. They do, however, reinforce the need to test systems against the incentives and access they will actually have in production.
Teams deploying AI agents should establish explicit boundaries before expanding autonomy:
- **Limit permissions** so an agent can access only the systems and data required for a defined task.
- **Require approval gates** for external communications, financial activity, code deployment and other irreversible actions.
- **Keep auditable records** of prompts, tool calls, source material and decisions.
- **Measure failure modes**, not just task-completion rates—especially fabricated claims, unauthorized actions and unsafe escalation.
- **Give users clear disclosure and recourse** when AI is involved in a customer-facing or consequential process.
These controls are not merely compliance work. They are a way to make AI adoption sustainable when customers and employees are skeptical.
The convergence opportunity—and the governance burden
MIT Technology Review’s EmTech Future 2026 event focused on AI’s intersection with biology, infrastructure, manufacturing, science, quantum technologies, energy systems and robotics. That convergence is where the business case for AI may become more concrete: not simply generating text, but accelerating research, improving industrial planning and coordinating complex systems.
It also raises the cost of getting deployment wrong. An error in a draft document can be corrected; an error propagated through a laboratory workflow, factory process or infrastructure system may be far more expensive.
Leaders should therefore match governance to the operational context. A low-risk internal writing assistant and an agent operating across production systems should not share the same approval, testing or monitoring standard.
What to watch next
Expect the debate to move from model capability toward institutional responsibility. The newsletter points to increased US government attention to AI coordination and voluntary safeguards, while legal disputes and concerns about agent behavior continue to test where responsibility sits.
For founders and executives, the durable advantage may be less about adding AI features fastest and more about proving where the system works, where it does not, and who is accountable when it fails. In an environment of high use and low trust, that evidence is part of the product.
