AI Ethics

TermCal uses artificial intelligence sparingly and transparently. We believe software should serve human judgment, not replace it. This page explains the principles that guide our use of AI and the broader technical and ethical context behind them.

1. Human agency is non-negotiable

Every AI-generated output in TermCal is a proposal, never a command. When our system extracts dates from a document or suggests a schedule adjustment, you remain the final decision-maker. We do not auto-commit events to your calendar, and we do not let algorithms override explicit human choices.

We design our features so that the cost of saying "no" is zero. If an AI suggestion is unhelpful, you can ignore it with a single click. There are no opaque optimization loops running in the background, no hidden engagement-maximising algorithms, and no dark patterns nudging you toward behaviour you did not choose.

2. The illusion of thinking

Recent research from Apple and MIT confirms what many practitioners already suspected: large language models do not reason or plan in any robust sense. In Apple's study, "The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity," frontier reasoning models faced complete accuracy collapse beyond certain problem complexities, failed to use explicit algorithms, and reasoned inconsistently across puzzles. Read the Apple paper →

Similarly, MIT researchers found in "Reasoning or Reciting?" that while language models appear to solve tasks, they often rely on narrow, non-transferable procedures memorised during training. When tested on counterfactual variants that deviate slightly from standard assumptions, performance degrades substantially. Read the MIT paper on arXiv →

These findings matter for scheduling software because calendars are systems of consequences. A missed deadline is not a harmless hallucination; it can affect contracts, employment, health, and relationships. That is why TermCal treats LLM outputs as probabilistic drafts to be verified by the user, not as authoritative decisions.

3. Gary Marcus: no substitute for well-specified algorithms

Cognitive scientist Gary Marcus has argued that large language models are "no substitute for good well-specified conventional algorithms." In his analysis of the Apple reasoning study, Marcus notes that LLMs fail at tasks—such as the Tower of Hanoi—where a simple deterministic algorithm would succeed every time. Read "A knockout blow for LLMs?" →

We agree. TermCal's core scheduling engine is built on explicit, verifiable logic: date arithmetic, recurrence rules, and conflict detection. Where we use AI (for example, document extraction), it is bounded to a single task with a human review step.

4. Yann LeCun: current architectures cannot learn to plan

Meta's Chief AI Scientist, Yann LeCun, has long maintained that current AI architectures are fundamentally insufficient for genuine reasoning and planning. In his position paper "A Path Towards Autonomous Machine Intelligence," LeCun asks: "How could machines learn to reason and plan?" and argues that autoregressive language models lack the world models and internal simulation required for reliable long-horizon decision-making. Read LeCun's paper on OpenReview →

We take this seriously. Until machine planning can be verified, audited, and overridden as easily as a spreadsheet formula, we will not delegate scheduling decisions to it.

5. Pope Leo XIV and "unthought thoughts"

In his 2026 Message for the 60th World Communications Day, Pope Leo XIV warned that uncritical reliance on artificial intelligence risks "turning people into passive consumers of unthought thoughts and anonymous products without ownership or love." He added that surrendering our mental capacities to machines would mean "burying the talents we have been given to grow as individuals in relation to God and others." Read the Vatican message →

The Pope's critique resonates with our design philosophy. Time is not a commodity to be optimised by an opaque algorithm; it is the medium through which human beings form commitments, relationships, and a sense of purpose. Software that manages time should therefore sharpen human judgment, not dull it.

6. The TESCREAL critique

Researchers Timnit Gebru and Émile P. Torres have traced the ideological roots of much AGI hype to what they call the TESCREAL bundle: a set of convergent worldviews that promise utopian futures through artificial general intelligence while downplaying present harms. In their paper for the DAIR Institute, they argue that this framework is "rooted in the Anglo-American eugenics tradition of the twentieth century" and that the systems built under its influence tend to "harm marginalized groups and centralize power, while using the language of 'safety' and 'benefiting humanity' to evade accountability." Read the TESCREAL paper →

We reject that trajectory. TermCal is not a stepping-stone toward AGI. We do not collect training data, we do not feed user behaviour into model improvement pipelines, and we do not make grandiose claims about the future of humanity in order to excuse present product flaws.

7. What we commit to

• Zero document retention. Documents uploaded for extraction are processed and immediately deleted. We do not archive your files.
• No training on customer data. Your calendars, commitments, and uploaded documents are never used to train or fine-tune AI models.
• Deterministic outputs where possible. Our scheduling logic uses explicit algorithms you could verify by hand. AI is used only for bounded extraction tasks.
• No black-box optimisation. We do not run engagement-maximising or behavioural-manipulation algorithms in the background.
• Human review by default. Every AI suggestion requires an explicit user confirmation before it affects your calendar.

8. Questions?

If you have concerns about how we use AI, or if you are a researcher interested in auditing our extraction pipeline, please contact us at hello@termcal.com.