AI in healthcare: one word, many realities
AI in healthcare is everywhere in the conversation. ChatGPT, augmented imaging, chatbots, flow prediction, automated scheduling… everything gets bundled into the same box. That is a dangerous shortcut, because these technologies do not work the same way, do not raise the same reliability questions, and do not apply to the same use cases.
For a hospital director, a nurse manager, or a CIO, understanding these differences is not an academic luxury: it is the condition for making sound technology choices, avoiding inflated promises, and keeping the decisions made with these tools safe.
Two broad families of AI
Connectionist AI, "the kind that learns"
This is the AI everyone talks about. It learns from large volumes of data. It excels at recognizing, classifying, translating, and predicting. It notably includes:
- Generative AI (ChatGPT, Claude, Gemini), which produces text, images, or code.
- Deep learning models used in medical imaging to support diagnosis.
- Predictive models that estimate patient flow in an ER.
Its strengths: raw power, the ability to process complex signals (image, language, time series).
Its limits: opacity, biases inherited from training data, hallucinations, dependence on massive volumes of labeled data.
Symbolic AI, "the kind that reasons"
Less in the spotlight, older, and yet at the heart of many critical applications. It relies on formalized rules and knowledge. Every decision can be explained, verified, and modified.
In healthcare, it is particularly well suited to anything that involves constrained decision-making: scheduling teams, checking regulatory compliance, supporting clinical decisions based on reference frameworks.
Its strengths: full transparency, traceability, no statistical bias, complete control by subject-matter experts.
Its limits: it does not "guess", you have to be able to formalize the rules and objectives.
Why Hopia chose symbolic AI for scheduling
Scheduling medical and paramedical teams is not image recognition or text generation. It is solving a constrained optimization problem: simultaneously handling hundreds of rules and producing a schedule that satisfies all of them while optimizing several objectives.
For this type of problem, symbolic AI is superior to learning-based AI for three reasons:
- Transparency, every assignment can be explained. A manager can ask "why is this person on this shift?" and get an answer grounded in the unit's rules.
- Compliance, a regulatory rule is a rule. It must not be "probably respected 95% of the time". It must be respected, period.
- No hidden bias, a model that learns from history risks reproducing past inequities. A symbolic approach starts from an explicit definition of fairness.
What this changes in practice
- For managers: they stay in control. The tool does not decide for them, it assists them by producing schedules that respect their rules.
- For caregivers: schedules are fair, explainable, and reproducible.
- For leadership: no black box, no dependence on foreign model providers, no legal risk tied to opaque automated decisions.
What Hopia is not
Let us be clear: Hopia does not do generative AI. We do not produce text, we do not answer patients, we do not write reports. We solve a precise algorithmic problem, optimized scheduling of human resources, with the technology best suited to that problem.
It is also a question of competitiveness: our algorithm is developed in France, by a French team, without dependence on foreign models. In a sector as critical as healthcare, that is a deliberate technology choice.