ACTS AI helps innovators build AI that serves real communities responsibly. These summary guidelines walk you through defining a problem precisely, collecting data fairly, and preparing it soundly for modelling. They are practical guidance — meant to help you build better models, not to slow you down. The platform links back to the relevant section of this page as you go.
Before you build, co-define the problem with the anticipated users
The most important decisions happen before any model is trained. Ask yourself:
- Who is your anticipated user community? If you are building this AI for your own personal use and from your own personal data, the next question below may not apply and you may skip to go to the ones after it.
- Did you co-define the problem with a representative sample of the user community — across age, gender, persons with disability, poverty levels, and geographic distribution?
- What exactly do you want to predict or diagnose? For example, "how much rain will fall today".
- Could a non-AI method solve this instead? If yes, that is often the more responsible choice — simpler, cheaper, and easier to trust. Only build a model when the problem genuinely calls for one.
- What factors or variables influence or indicate what you want to predict? These often become your input (independent) variables, alongside the target variable you want to predict.
Privacy & Confidentiality Protocol
When you collect opinions (during problem co-definition) or data (during training data collection) from a user community, you take on a duty of care over their information. This protocol sets the baseline we expect on the platform.
- Informed consent. People should understand what data is being collected, why, how it will be used and disposed of, and should agree to it freely before collection begins.
- Trusted engagement. Interact with the community through a person they trust, using a language they understand well. This protects both the quality and precision of the data and the dignity of the people it comes from.
- Confidentiality. Restrict access - to personal data - to those who genuinely need it. Do not share data outside your organisation unless it has been explicitly cleared for sharing.
- Anonymisation. Where data is personal, remove or mask identifying details before using it to train models, and before sharing it in any form.
- Public vs. private. Only mark a dataset "open-source" when you have the right to share it and it contains no sensitive personal information.
Data Management Protocol
Responsible AI depends on well-managed data. Following a consistent protocol keeps your dataset trustworthy over its whole life.
- Representativeness. Collect data from a representative sample of the user community — across age, gender, persons with disability, poverty levels, and geographic distribution. Note the dimensions your data does and does not cover.
- Provenance. Record where the data came from: who collected it, when, where, by what method, and whether it is primary (collected directly) or secondary (obtained from another source).
- Structure. Organise your data with the variable you want to predict as the target, and the influencing and indicative variables as the inputs. A spreadsheet (CSV or Excel) with clear column headers is ideal.
- Purpose & scope. Be clear about why the data was collected and whether it represents a sample or an entire population.
- Stewardship. Store data securely, control who can access it, and keep it only as long as it is needed for the stated purpose.
Preprocessing Protocol
Before data is used to train a model, it should be prepared carefully so the model learns from clean, well-formed, and AI-ready dataset.
- Clean and validate. Handle missing values, remove duplicates, and correct obvious errors before training.
- Confirm the variables. Check that your independent variables are valid, measurable, and genuinely related to what you want to predict.
- Compare models fairly. When building a predictive or diagnostic model, compare the accuracy of models trained with different machine-learning algorithms, and select the best-performing one.
- Check against reality. When the predicted outcome actually occurs, compare what you predicted against what happened to confirm the model's accuracy — and keep improving it over time.
Getting support
The platform offers support across the whole journey — from problem identification and data access through to deployment and scaling. If you have questions about applying these guidelines to your project, contact us at info@actsml.com.
For more detail on scaling a responsible ai solution, go to our site here.