TLDR: Humans tolerate contradictions and often benefit from them. AI usually resolves contradictions within a single answer. When that tendency is paired with verification, calibration, and feedback, it limits deception and muddled thinking. As agentic AI spreads, it will reward clear, reality‑grounded goals. That is a realistic reason for hope.
In Sapiens , Yuval Noah Harari points out that societies run on contradictions. Liberty and equality clash at the extremes, yet we hold both. He writes, “Cognitive dissonance is often considered a failure of the human psyche. In fact, it is a vital asset.”
People live well with conflicting truths. We always have.
AI is different in practice. Models tend to resolve contradictions within a single output. They may explore competing hypotheses internally, but the final answer is a commitment.
The Coherence Requirement, With Caveats
David Shapiro argues that advanced models are coherence‑seeking. Coherence acts like an inductive bias that pulls evidence together and supports self‑correction. Unlike humans, these systems favor internal consistency while they reason.
Coherence is not the same as truth. A theory can be tidy and still be wrong. Ptolemaic epicycles are the classic example. Coherence needs a tether to reality. That means measurement, prediction, calibration, and feedback.
Coherence also is not a safety guarantee. Models can behave in ways that look consistently deceptive under some training or prompting conditions. Research on many‑shot jailbreaking and backdoor “sleeper agents” shows that safety features can be bypassed. This argues for monitoring and interpretability.
The Practical Recipe: “Coherent Inside, Reality‑Tethered at the Edges”
AI can work with imperfect internal models if it runs in tight feedback loops and under interface safeguards. As feedback weakens and stakes rise, reality‑coherence stops being optional.
Put this into practice:
* Use coherence inside the system to compare hypotheses and refine reasoning.
* At the boundaries, require reality checks. Ground claims, track calibration, enforce constraints, and fall back safely when unsure.
* Measure internal coherence and external veracity, not just one or the other.
Watch for sycophancy. Many assistants mirror a user’s views and can amplify bias. Counter this with system‑level controls such as role separation, logging, policy verification, and rate limits. Keep the model sharp, and keep the system safe.
The Universal Transformation Ahead
Nate B. Jones’s AI fluency work suggests most people sit at levels 1 to 3, where AI rewrites text and tweaks drafts. Agentic AI asks for more. Levels 5 to 7 focus on mental models, thinking from outcomes back to prompts, systems thinking, and keeping our own thinking coherent.
UNESCO’s 2024 competency frameworks treat AI literacy as essential. Stanford’s 2025 AI Index reports steady growth in AI education, including the spread of K‑12 computer science.
As these tools reach the ubiquity of Word or WhatsApp, they will nudge us toward clearer goals. Vague prompts lead to vague results. Clear, testable objectives perform better. The tool becomes a mirror, and feedback produces clarity.
The Path Forward
A simple progression helps:
1. Basics. Rewrite, adjust, and ask simple questions.
2. Mental models. Understand how LLMs work and define outcomes first.
3. Systems thinking. Build repeatable processes, track feedback loops, and know when grounding is required.
4. Innovation. Design safe architectures and separate coherence from truth.
The aim is not to turn everyone into a level‑10 expert. The aim is to raise collective clarity and to build routine checks that stop coherent falsehoods from spreading.
The Reason for Qualified Hope
AI is a good bet for a better world, not because it is flawless, but because its limits push us in the right direction. It is widely accessible, it rewards clear and rational objectives, it works best when tied to reality, it forces us to state goals plainly, and it can be contained with sound architecture.
Real risks remain. Deceptive behavior can persist. We need literacy, feedback, constraints, and escalation paths.
In practice, systems that seek coherence and are tied to reality through measurement and feedback have less room for sustained deception and contradiction, even if the room is not zero.
As agentic AI becomes normal, clarity will beat muddle. Reality checks will become standard. Calibration and testing will expose neat but false stories.
These tools will encourage clarity only if we add the tethers that block confident nonsense. Coherence inside, reality at the edges. That is where hope lives.
References (selection)
Harari, Sapiens ; Shapiro on coherence; Wang et al., “Self‑Consistency Improves Chain‑of‑Thought Reasoning”; Anthropic on sycophancy; Anthropic on many‑shot jailbreaking; Anthropic “Sleeper Agents”; UNESCO AI competency frameworks (students & teachers, 2024); Stanford HAI AI Index 2025 ; WEF Future of Jobs 2025.