EXIN BCS Generative Artificial Intelligence Award (AIGAIA) Exam Syllabus

• a. Artificial intelligence (AI) – Intelligence demonstrated by machines, in contrast to the natural intelligence displayed by humans and other animals.
• b. Generative artificial intelligence – Deep learning models that can generate high-quality text, images and other content based on the data they were trained on.
• c. Large language models (LLMs) – Deep learning algorithms that can recognize, summarize, translate, predict, and generate content using very large datasets.
• d. Natural language processing (NLP) – The ability of a computer program to understand human language as it is spoken and written.
• e. Prompts – The inputs or queries that a user or a program gives to an LLM AI, in order to elicit a specific response from the model.
• f. Completion – The output or result generated by AI after processing and understanding the provided prompt.

- Guidance

• Candidates will be able to recognize and recall the definitions of key generative AI terminology as listed.

• a. For personal or organizational use
• b. Respond to queries, improving search
• c. Content creation
• d. Summarize documents
• e. Text to image, image to text
• f. Following instructions
• g. Writing computer programs

- Guidance

• Generative AI is used in an enormous variety of tasks in social and work environments with varying levels of success, risk and responsibility.
• Candidates should be able to recognize and describe the use of generative AI in contexts such as answering simple text-based questions, creating reports, summarizing large volumes of text, writing accessibility text to describe images or writing code to program a computer.

• a. Machine learning – The study of computer algorithms that allow computer programs to automatically improve through experience.
• b. Deep learning – A multi-layered neural network.
• c. Stages of the machine learning process:
  - Analyze the problem
  - Data selection
  - Data pre-processing
  - Data visualization
  - Select a machine learning model (algorithm)
  1. Train the model
  2. Test the model
  3. Repeat (learning from experience to improve results)
  - Review

- Guidance

• The machine learning process allows us to define the solution based on the problem that has been identified through the process of data selection, pre-processing, visualization and testing of data with specific algorithms.
• There is no de facto method within machine learning, learning through experience is vitally important to generative AI, to help improve the quality and relevance of the output. Testing involves creating the correct test data, creating bodies of data to learn from and parameters for what you wish to test.

• a. Testing
• b. Training
• c. Reinforcement learning
• d. Reinforcement learning from human feedback (RLHF)
• e. Inferencing

- Guidance

• Candidates should be able to describe each of the stages of the generative A Iprocess as listed. The model is firstly trained using vast data sets, then tested using controlled, unseen data. Then, reinforcement learning takes place, where AI learns from the perceived quality of its output or response and uses this to improve its output in future. This takes place in RLHF, where human operators pose thousands of prompts to the AI model, checking the response, then ‘rewarding’ the AI model for correct responses.
• Inferencing is when a trained and tested AI model is fed new data, and prompted to generate a response, such as a prediction or recommendation.

• a. Training data including pre-training data
• b. Test data

- Guidance

• In generative AI, good quality training and testing data is incredibly valuable. The training data is used to train the model, while the testing data is used to evaluate its accuracy.
• Training data is used to feed the AI model enormous banks of information, which it then uses to construct a response to a prompt. Pre-training data is the first batch of data which is fed to the model, without any refinement or finetuning. Training data is the term used to describe the data used thereafter, which is more focused or specific. The quality of the data used for training has a direct impact on the quality of the generated output.
• Test data is unseen data – data which has not been used in any training capacity – which is used to assess the performance and output of the AI model.

• a. To make predictions
• b. Required for long responses

- Guidance

• Transformers help to provide more accurate predictions about the next most likely word, phrase, sentence, and even paragraphs in response to a prompt. A transformer provides the capability for lengthy responses – running into thousands of words – although those responses might not be accurate.

• a. Supervised fine tuning (SFT)
• b. Reinforcement learning from human feedback (RLHF)

- Guidance

• Candidates should understand the role of both RLHF and SFT in providing feedback on the responses of generative AI. In SFT, the desired response to a prompt is created by a human and this response is used as training data. In RLHF, human operators pose thousands of prompts and carefully check the response, ‘rewarding’ the chatbot for correct responses.
• This is an ongoing process – constant fine tuning. This is why we see constant improvement.

Prompting generative AI - 10%

• a. To request an output
• b. Prompt engineering

- Guidance

• A prompt is the instruction given to the generative AI model by the user. It powers the transformer – which looks at the prompt, at the training data, and at what it is generating at the same time. This is why a slight change to the prompt or how it is worded can affect the output.
• Prompt engineering is the art of altering and refining prompts, to reach a desired, or better-quality output.

• a. Zero-shot, one shot, few shot
• b. Character
• c. Chain of thought

- Guidance

• Zero-shot prompts are short, basic prompts with no additional instruction or context. Character prompts are when AI is asked to create the output in a particular tone or style, based on characteristics such as a given character, time period, or geographical location. Chain of thought prompts are more complex problems, which require multi-level reasoning in order to construct a response.
• The more examples you include, the better the output.

• a. Human verification
• b. Fact checking
• c. Checking cited sources

- Guidance

• Generative AI is capable of “hallucinations”. This is when an output presents false or misleading information as fact, often the result of an ambiguous prompt. Examples of this are citing false sources, biased information, or false positives.
• This creates a need for human fact verification and fact checking, to ensure that any AI generated output which is being used or shared is correct and fit for purpose.

• a. Subject matter experts (SMEs)
• b. Reword the prompt to compare output

- Guidance

• Actions can be taken to assess the validity of generative AI output. Reviews by SMEs can be used to identify errors, bias or false information. Prompt engineering can also be used in validation. By giving the same instruction, worded in a different way, humans can assess if the generated outputs match and are consistent, allowing any discrepancies to be investigated. This method would still require human input.

Ethical and legal concerns - 25%

• a. Data sources:
  - Malicious
  - Commercially sensitive
• b. Bias
• c. Inaccuracies and false information

- Guidance

• In the development of generative AI, consideration must be given to the potential ethical concerns of the data being used for training, and the output this creates.
• Candidates must consider the sources of data being used for training and testing and their reliability. For example, if data comes from a source with a particular political or moral stance, it is likely to contain bias and false or misleading information. Equally, commercially sensitive or personal data should not be used to train AI, and this could contain information which poses a risk to individuals or organizations if shared.
• Using ethically questionable data to train and test AI could lead to poor output, containing bias or false information.
• Candidates should be able to identify simple opportunities for AI in an organization, such as an opportunity to automate a process, or minimize the human input into a repetitive task.

• a. Copyright
• b. Plagiarism
• c. Data storage and use
• d. Data security and privacy

- Guidance

• Candidates should be aware of both the legal and regulatory items to consider when developing and using generative AI. In developing generative AI, the use and storage of data must be compliant with relevant legislation, such as UK Data Protection Act, UK GDPR, and Privacy and Electronic Communications Regulations (PECR). If working outside of the UK, consideration must be given to the specific legislation relevant to the country of operation. In using AI, candidates should consider the input and output of the AI model, and always check the output for use of copyrighted content. The data used in the prompt should also be considered – as data entered into a generative AI model cannot be guaranteed to be secure. Private, legally protected or commercially sensitive data should not be used in prompts. Organizational guidelines and policies should also be adhered to.

• a. Reverse search the output
• b. Prompt quality
• c. Keep humans involved

- Guidance

• Steps can be taken to minimize the risks presented by generative AI. Candidates should be able to explain and suggest suitable mitigations. Reverse-searching the output of the AI model can be used to identify if the content already exists somewhere online, this can be helpful in identifying copyrighted or plagiarized content. Improving the quality of the prompt input can help to avoid hallucinations and can significantly improve the quality and relevance of the output. Human input throughout the use of generative AI is key to mitigating and minimizing risk, as common sense and expertise can be applied to the prompt, the output and the application or implementation of it.