MM1041 Chap.9 Risks, Bias and the Limits of Machine Learning
Risks, Bias and the Limits of Machine Learning
Five risks named in the first lecture
The lecture lists five risks and limits. Hidden biases: a system that learns which job applicants to interview from past recruiters' decisions can learn to perpetuate their racial, gender or ethnic biases. Lack of verifiability: deep neural networks often cannot explain how they work.
Lack of data: new problems, and problems without useful data behind them, remain work for human specialists. Unintended consequences: a machine has no grasp of what its advice will lead to.
And errors, such as spam filters blocking important emails.
Errors you have met
The lecture asks students about their own experience of AI errors and lists familiar ones: machine translation that corrupts meaning, autocorrect choosing the wrong word, biometric systems misrecognising people, transcription software missing speech, and a self-driving car hitting a wall.
Small individually, such errors become costly when repeated at business scale.
Five safeguards
The lecture's response is practical. Control what users feed the system and limit learning to verified inputs. Test algorithms for bias by race, gender, age and similar attributes. Analyse explicitly how the software can fail and provide a safety mechanism for each failure.
Keep a less smart backup product or service. Decide in advance what the firm will tell journalists if a failure becomes public.
Bias in a high-profile launch
The lecture's bias case is an image generator whose launch in February 2024 produced inaccurate and offensive images, drew criticism from prominent public figures and led to an apology from the company.
It shows that bias is not only a question of which groups appear in training data; any systematic error in what a system produces can harm users and the firm's reputation.
Deepfakes in Hong Kong
The lecture brings the topic home with Hong Kong commentary on a wave of deepfake crime, in which anyone with an image can become material for fraud, and with a deepfake of a well-known entrepreneur promoting a cryptocurrency scam.
As generative tools improve, the burden shifts to verifying who and what is real.
Why this chapter is examinable
The course's fourth learning outcome is awareness of contemporary ethical issues and the impact of AIDA on business and society, and the exam rubric grades it directly. A strong answer names a specific risk and a specific safeguard for the case in front of it.
What this chapter covers
- 01
Hidden bias from historical decisions
- 02
Unverifiable deep learning output
- 03
Problems without data
- 04
Unintended consequences of machine advice
- 05
Everyday AI errors
- 06
Five safeguards
- 07
Deepfakes and fraud
Worked example · free
Pair risks with safeguards for a lending model
- 2Identify and explain the first risk.
- 2Identify and explain the second risk.
- 2Give a safeguard for each.
Key terms
- Hidden Bias
- Unfair patterns a model learns from historical data without being instructed to discriminate.
- Verifiability
- The ability to check and explain how a system reached its output.
- Unintended Consequence
- A harmful result of machine advice that the system could not understand or foresee.
- Deepfake
- Synthetic video, image or audio that convincingly imitates a real person.
- Backup System
- A simpler product or process kept ready to take over when an AI system fails.
- Communication Plan
- A prepared approach for explaining an AI failure to the media and the public.
- Verified Input
- Data checked for accuracy and origin before a system is allowed to learn from it.
- Misrecognition
- An error in which a biometric or vision system identifies a person or object wrongly.
Risks, Bias and the Limits of Machine Learning FAQ
Why can an AI system be biased if nobody programmed it to be?
Because it learns from data, and data about past human decisions carries the patterns of those decisions. The lecture's example is a hiring system trained on recruiters' earlier choices, which can learn and repeat their racial, gender or ethnic biases. The bias enters through the training data, not through any explicit rule.
What should a company do before launching a machine learning system?
The lecture lists five steps: limit learning to verified inputs, check the algorithm for common biases, analyse how the software can fail and add a safety mechanism for each failure, keep a less smart backup, and prepare a media communication plan. Together they cover prevention, detection, fallback and reputation.
Why are deepfakes treated as a business risk and not only a crime story?
Because they attack the trust businesses rely on. A convincing fake video or voice can impersonate an executive, endorse a fraudulent product or trick staff into a payment. The lecture's Hong Kong examples show the threat is local and current, so firms need verification steps that do not rely on recognising a face or a voice.
Exam move
Memorise the five risks and the five safeguards as two lists, then practise linking them on three different cases, one in hiring, one in finance and one in customer service. For each, write one sentence naming the risk and one naming the safeguard; that pair is the unit an ethics answer is marked on.
Working through Risks, Bias and the Limits of Machine Learning in MM1041? Sia is AskSia’s AI Information Technology tutor — ask any MM1041 Risks, Bias and the Limits of Machine Learning question and get a clear, step-by-step explanation grounded in how MM1041 is taught and assessed. Read this chapter free, then take your hardest questions to Sia.
