Find out what you should and shouldn't do when it comes to using Generative AI:
Do
- Define your objectives, articulating Generative AI-related goals and objectives.
- Understand the capabilities and limitations of different Generative AI models and implementations.
- Ensure you have appropriate policies and guidelines in place as to the use of Generative AI.
- Ensure you, and those using generative AI in your organisation, have a base level understanding of how the technology works as a minimum and ensure staff using Generative AI systems are adequately trained.
- Start with small projects or proof-of-concepts that are not business critical. Learn from these, about potential impacts and feasibility, before scaling up.
- Prepare your data, focussing on accuracy, hygiene, quality and diversity
- Continuously monitor and evaluate Generative AI outputs against your objectives and standards and against expected outputs.
- Include humans in the loop as necessary. Review and challenge outputs with professional scepticism, avoiding automation bias.
- Create detailed prompts – explain exactly what you want Generative AI to do and give it a good example of how to do something.
- Make it clear when Generative AI has been used in producing content;
- Consider a scientific approach to using generative AI, where rigour and accuracy are important – apply a tight framework and repeat prompts to reduce the risk of anomalous responses.
- Be curious and experiment! But be responsible too.
Don't
- Ignore ethical aspects: develop guidelines for responsible use of generative AI.
- Abdicate responsibility: generative AI models and outputs require human oversight.
- Overestimate generative AI models: they are tools to complement human expertise.
- Underestimate the hardware or infrastructure requirements of generative AI. It requires significant volumes of good quality data.
- Disregard user feedback: use it to identify and address issues, as well as improve models.
- Fall foul of data protection, privacy and intellectual property considerations: keep client and confidential internal data off public generative AI tools, and make sure you have permission to train a model using someone else’s IP or data.
- Ignore potential inconsistencies and inaccuracies in generative AI outputs.
- Forget that the implementation of new technology requires cultural change: supporting those with legitimate concerns about AI is crucial to its successful adoption.
As considerations vary across sectors, organisations, use cases and data, these do’s and don’ts should be evaluated and adapted to reflect specific needs.