Responsible AI requires dialogue, not a checklist.
About the ethics behind AI
By Peter Reddering, Domain Lead Data Strategy @ Cmotions, Katinka Knops, – Ethicus @ Digital Ethics – advice on ethics and digitalization & Jurriaan Nagelkerke, Data Science & AI Tech Lead @ Cmotions
Developments around AI are moving at a rapid pace and the promises are immense. Data is indeed ‘the new gold’, so what could be more logical than having data scientists dig into the data and start mining? Yet the real questions always seem to be not only technological, but also ethical: what do we consider desirable, who will experience the consequences, and which values are important?
In this article, we show that this ethical perspective is not a side issue, but a thread that runs through the development of every AI application. Technology is not neutral and when developing it, you constantly make choices. Therefore, you weave that thread from the very beginning. That starts with the right dialogue.
The choices you make no matter what
Anyone who develops an AI application makes decisions from day one that determine how the application behaves and who experiences the consequences of that, whether intended or unintended. Think of:
- which data you use to train the AI model or to fine-tune your prompts – garbage in, gospel outTo update an old wisdom, just to put it in perspective;
- which data does the model rely on when used in daily practice to make decisions – behavioral data or special personal data without the knowledge of the individuals involved?;
- what kind of model you use: self-trained, fine-tuned, or a foundation model that you ‘request’;
- where models, imports and exports are stored – and what monitoring you set up to see if things are going well.
- how the user can (co-)work with the model.
Each of these choices sounds technical, but at its core it is also ethical: with each choice you weigh values and decide which values take priority. Employees have to work with the results, customers have to be able to interact with a chatbot, decisions are made that have an impact on citizens. Our practice shows that the most important choices are not about the technology itself, but about the organization: what goals are you aiming for, which human considerations do you automate, and which values are so important that you want to ensure they are included in your design? Scoring with AI therefore does not start with the technology. It starts with the question of what you as an organization want to stand for.
From the practice — No more calling
A government organization in the healthcare sector assesses applications for care or care budgets. An important step in the process: a phone call with the applicant to check that everything is complete and that the client understands the implications of the application on their personal situation. To save scarce resources, it is being investigated whether this can be automated. Smart technology will soon check the completeness and request missing information digitally. Efficient! But… with this, the only personal contact moment may also disappear. Is that desirable – for the organization, the employees, and the client? Or is the emphasis on efficiency quietly sacrificing other values here?
Neutral?
Anyone who thinks they can escape ethical considerations by relying on the neutrality of technology is deluded. An algorithm consistently and without complaint calculates results, and therefore the results appear objective. But it is calculated based on data that people have collected, with criteria that people have devised, towards goals that people have chosen. In short, there is surprisingly much human — and therefore surprisingly much morality — in that ‘neutral’ machine. Train a model on the past and it consistently reproduces that past, including the skewed perspective that was present in it. An algorithm that labels a particular postcode as ‘risk’ is not visionary. It is a mirror with a selective memory.
From practice — the diligent parrot
A youth care organization is developing predictive models to estimate risks of problematic behavior, trained on historical records. The model neatly distinguishes between high and low risk. Until in-depth analysis reveals that young people with a certain cultural background have simply been reported and followed more frequently in the past – possibly due to differences in interpretation or biases. The model dutifully learns: that group = higher risk. Even without ancestry as a variable, postcode and school type sneak in as indirect indicators. Thus the algorithm becomes a diligent parrot of old patterns, with self-fulfilling prophecy as a risk. Technically, it does exactly what is asked of it. That is precisely the problem.
The checklist desire
So if we can’t simply rely on the neutrality of the technology, how can we make the right choices in design? Perhaps a list of criteria? In our conversations and sessions with clients, we keep seeing this need for checklists and frameworks. Understandable: A checked list gives a sense of control. ‘Look, I’ve gone through all the criteria, so I can justify myself.’ The problem is that ethics cannot be reduced to neat categories: the truly relevant questions are rarely included on a standard list, because they depend on the specific application and context.
What is innocent in one situation (an algorithm that suggests outfits based on the contents of your wardrobe) is not the same in the other (an algorithm that decides on your loan based on gender, age, and postcode). The ethical perspective is therefore not a form that you fill out once and archive it; it is a continuous dialogue about values and the consequences of your choices: who benefits from this application, who might be harmed by it, and can we accept those consequences together? In ethical issues, there is rarely a black-and-white ‘good’ or ‘bad’; it is about a balanced argument that supports choices. It is precisely in that balance – and the transparency thereof – that careful accountability lies. Not in a standard form filled out for 100%.
But shouldn’t it be allowed by law?
Those who prefer to ignore the conversation about values have one more trump card: the law. ‘We act 100% ethically because we remain completely within the legal framework.’ A reassuring thought, but ethics and legislation are not the same thing. In fact, ethics precedes legislation. In the law, we establish what we have agreed upon as a society after a social and moral debate. By doing so, legislation is by definition behind the reality, especially in a technology that changes much faster than an average legislative process! Moreover, something that is legally permitted cannot necessarily be a good idea that fits the values of your organization and stakeholders.
How then?
The common thread is clear: the ethical perspective is present in every step—in the choices you make, in the technology you build, and in how you justify your actions. Deploying AI responsibly therefore begins not with an algorithm, but with the right people at the table and a few “awkward questions.” Not much glamour, but a lot of impact. This naturally raises practical questions: how do you organize that conversation and who do you involve? How do you document the tradeoffs? And how do you structurally embed ethics within the organization so that it doesn’t remain a one-time exercise but becomes a constant part of how you work with AI—without it becoming, you guessed it, a checklist? In the follow-up to this article, we will further explore these questions!
Learn more
Cmotions helps organizations use AI smartly and responsibly in their development process. Do you have questions about agentic working, governance around AI tooling, or scaling up Claude Code within your team? Feel free to contact us.
