Can an Algorithm Be Fair | Academic English

Start talking

Start talking: first response

A hiring tool is correct in nine cases out of ten. Is that enough information to recommend it?

Take it further

Which question would you ask before hearing another accuracy statistic?

Clip A

Watch for: How does one software failure travel between very different projects?

Academic Vocabulary

training set

Examples used to fit a machine-learning model.

a representative training set

The team needs a training set that includes the people who will use the system.

underrepresentation

Insufficient inclusion of a group relative to the purpose of a sample or dataset.

underrepresentation in the data

Underrepresentation in the data may conceal poor performance for some users.

computer vision

Methods that enable computers to extract information from images or video.

a computer vision system

The computer vision system detected a face but did not identify the person.

algorithmic bias

Systematic patterns in an automated process that can produce unfair or skewed outcomes.

investigate algorithmic bias

We should investigate algorithmic bias before expanding the service.

Original source

How I'm fighting bias in algorithms | Joy Buolamwini · TED · 00:12–04:04 · Source and credit

Clip A quiz: 6 questions

Choose one answer for each question, then read the feedback.

1. What difficulty appears in Buolamwini’s demonstrations?

The software fails to detect her face.The software detects her face but changes her name.The software identifies everyone but cannot store images.

2. What did the white mask reveal in the Aspire Mirror project?

The mirror needed a spoken password.Recognition was deliberately disabled for every researcher.Detection depended on how the face appeared to the system.

3. Why is the Hong Kong example important to her argument?

The robot was designed to test national language skills.Reused software carried the same exclusion elsewhere.A different country caused the software to lose its training data.

4. What does a training set contribute to the explanation?

A list of moral rules automatically enforced by the computer.A record of only the errors made after deployment.Examples from which a system learns patterns.

5. Why can a narrow training set cause detection failures?

A smaller dataset always runs more slowly.Any variation makes image analysis impossible.Some users differ from the examples the system learned.

6. What opportunity does the speaker identify at the end of this excerpt?

The existing software will correct every error unaided.Computer vision must be abandoned in all settings.People can create more inclusive training sets.

Grammar in a minute: Concession with despite

Despite takes a noun phrase or an -ing clause. Use although or even if before a complete finite clause.

You say: Acknowledge a system’s strength before naming a limitation.

One possible response

Despite reducing processing time, the system still needs evidence of reliable performance across groups.

Useful English

Useful English: speak and compare

Say your response before opening a model.

Average performance can conceal …

Question an aggregate result.

You say: Respond to a vendor who offers one overall accuracy score.

One possible response

Average performance can conceal uneven errors. I would like to see which applicants the system misses and how often.

The failure may lie in … rather than …

Distinguish a plausible mechanism from a superficial explanation.

You say: Explain why a failed face detector is not evidence that someone has an unusual face.

One possible response

The failure may lie in the training coverage rather than in the person. We should investigate how the system was built.

Critical Thinking: 3 questions

Choose one question, then follow up, challenge and revise. Use the others for another angle.

Would a more diverse training set be sufficient to make a hiring system fair?
Help me start
Consider the target, the data and the decision made with the output.
One possible response
It could address an important source of error, but it would not settle what the system rewards or how scores are used. A well-represented dataset can still encode an unsuitable definition of success.
Take it further: What unfairness could remain with equal detection accuracy?
C2 challenge
Separate technical performance from a defensible institutional objective.
When a shared software component fails, who should be accountable?
Help me start
Compare the developer, the purchaser and the deploying organisation.
One possible response
Responsibility can be shared without becoming vague. Developers should disclose limitations, while deployers should establish suitability for their users. The organisation making consequential decisions cannot treat an external component as an excuse to stop asking questions.
Take it further: Who should respond first to an excluded user?
C2 challenge
Allocate one concrete obligation to each actor without duplicating every obligation.
Is the white-mask demonstration persuasive evidence of the full scale of the problem?
Help me start
Separate showing a failure from estimating its prevalence.
One possible response
It makes the failure visible and challenges an assumption of universal performance. It does not by itself estimate how often the problem occurs across systems or groups. I would use it to motivate a wider, structured evaluation.
Take it further: What would that evaluation need to sample?
C2 challenge
Defend the demonstration’s value while limiting the inference drawn from it.

Clip B

O’Neil is concluding earlier examples of hiring, teaching and criminal-risk scores. Her audit asks whether the data, the definition of success, the errors and the system’s wider effects deserve confidence.

Watch for: Which parts of a model must be questioned beyond its overall accuracy?

Academic Vocabulary

audit

A systematic examination of evidence, processes and outcomes against explicit criteria.

conduct an algorithmic audit

The panel will conduct an algorithmic audit before recommending wider use.

proxy

An observable measure used to stand in for something harder to measure directly.

a proxy for ability

Past promotion may be a poor proxy for ability when opportunities were unequal.

false negative

A result that incorrectly classifies a case as lacking the condition being tested.

a costly false negative

Rejecting a qualified candidate would be a costly false negative in our fictional screening system.

feedback loop

A process in which an outcome changes the conditions that produce future outcomes.

reinforce a feedback loop

Using only past hires to train the next model could reinforce a feedback loop.

Original source

The era of blind faith in big data must end | Cathy O'Neil · TED · 09:59–12:33 · Source and credit

Clip B quiz: 6 questions

Choose one answer for each question, then read the feedback.

1. Why does O’Neil trace algorithmic bias back to human choices?

Bias arises only when someone changes the final score.People select data and define success.People must manually approve every calculation.

2. What does her data-integrity example ask us to distinguish?

Present data from all historical information.Recorded arrests from underlying behaviour.A numerical record from any usable evidence.

3. Why might staying and being promoted be an imperfect hiring target?

It can reflect organisational support as well as ability.Promotion is never relevant to employment.A four-year record necessarily contains too little data.

4. What purpose does the blind-audition example serve here?

It replaces the need to define a good performance.It illustrates deciding what evidence is relevant.It proves every occupation can be assessed without context.

5. Which questions extend accuracy into an assessment of errors?

How quickly, in which language, and with what logo?How much data, regardless of who is affected?How often, for whom, and at what cost?

6. Why does the final audit stage look beyond immediate predictions?

Feedback loops occur only outside social systems.Decisions can reshape future conditions and data.Long-term effects replace any need for present accuracy.

Grammar in a minute: A limit on a claim

Insofar as means to the extent that. It qualifies the scope of a claim rather than offering an unlimited concession.

You say: Limit a claim about a dataset’s usefulness.

One possible response

The dataset is useful insofar as it reflects the people and decisions we actually need to evaluate.

Useful English

Useful English: speak and compare

Say your response before opening a model.

What counts as success here?

Expose an implicit objective.

You say: Challenge a system rewarded for hiring people who stay longest.

One possible response

What counts as success here? Retention may matter, but it could also reflect who receives support after being hired.

Who bears the cost of that error?

Bring consequences into evaluation.

You say: Discuss a screening system that incorrectly rejects some qualified candidates.

One possible response

Who bears the cost of that error? The institution saves interview time, while excluded applicants lose an opportunity they may never know they deserved.

Critical Thinking: 3 questions

Choose one question, then follow up, challenge and revise. Use the others for another angle.

Should a hiring model prioritise avoiding unsuitable hires or avoiding the rejection of qualified people?
Help me start
Consider the next stage, the stakes and opportunities to correct errors.
One possible response
I would consider the whole process. If a human interview can identify unsuitable applicants later, early screening should be cautious about rejection. In a role with immediate high consequences, different safeguards may be justified.
Take it further: What if rejected candidates can appeal?
C2 challenge
State which change in error costs would reverse your preferred threshold.
Can an audit settle a disagreement over fairness?
Help me start
Separate empirical questions from competing values.
One possible response
An audit can reveal who is affected and whether claims hold. It may not resolve disagreement about which outcomes matter most. I would make those value choices explicit instead of presenting a technical score as their substitute.
Take it further: Who should participate in defining the criteria?
C2 challenge
Explain one possible conflict between two reasonable fairness objectives.
Would publishing a model’s code make it accountable?
Help me start
Consider access, explanations, recourse and institutional responsibility.
One possible response
Code access may help specialists investigate, but affected people also need understandable reasons and a route to challenge decisions. Accountability requires someone with authority to respond; disclosure alone may leave the practical power unchanged.
Take it further: What information should an unsuccessful applicant receive?
C2 challenge
Distinguish inspectability from the ability to obtain a remedy.

Mastery

Mastery: 6 questions

Original cases. Choose, then explain one answer aloud.

1. Fictional model: overall accuracy is high, but one group experiences twice as many mistaken rejections. What follows?

The difference proves a deliberate intention to discriminate.Average accuracy leaves a consequential distribution hidden.The model must be inaccurate for every individual.

2. A training set includes all groups equally, but rewards past promotion in a company with unequal support. What remains uncertain?

Whether the target represents ability fairly.Whether the dataset contains enough demographic categories.Whether numerical targets can be calculated at all.

3. A system learns only from people it previously hired. What risk needs attention?

Its decisions shape the evidence used to justify future decisions.Its training becomes independent of institutional choices.Its accuracy can no longer be measured in any way.

4. A vendor offers a larger dataset as its only response to unequal errors. Which next question is strongest?

Why is every small dataset automatically fairer?Can the same accuracy number be shown in a larger font?Which missing cases and error mechanisms would the extra data address?

5. A pilot lets every rejected applicant obtain a fresh independent assessment. What does this add?

Proof that the model’s original labels were unbiased.A reason to stop measuring unequal error rates.A route to correct individual decisions.

6. A panel disagrees over acceptable error costs after receiving reliable group statistics. What kind of question remains?

A question that more training data must automatically settle.A value choice about consequences and responsibilities.Only a missing arithmetic calculation.

Grammar in a minute: A cautious passive conclusion

Cannot be assumed to + base verb rejects an unsupported inference. It does not claim that the inference is certainly false.

You say: Qualify a claim about automation and fairness.

One possible response

An automated decision cannot be assumed to be impartial simply because no person pressed the reject button.

The hiring committee

Integrated spoken task: decide and revise

Fictional Northbridge Services wants to introduce a screening model. A vendor reports faster processing and strong overall accuracy.

Choose a full launch, a limited pilot or rejection. Name the evidence that matters and one condition for changing your decision.

Open only your own role. Share its information by speaking.

Recruitment director

Reduce delays while selecting people who can do the job.

Current applications take six weeks to process. The vendor defines success as staying for three years and receiving a promotion.

Applicant advocate

Protect access, intelligible decisions and a way to contest errors.

The firm’s mentoring programme previously reached only staff based at headquarters, although the model uses promotion records from every location.

Give an initial recommendation and its strongest reason.
Exchange private information; challenge one assumption and answer the objection.
Hear the new information, then revise or defend the decision with a stated reason.

New information: revise your decision

A new pilot comparison finds substantially more mistaken rejections at regional offices, despite similar independently assessed work samples.

Reconsider both the target and deployment. What must change, and what work can continue?

A strong response

Use a relevant idea from each clip.
Answer the other person’s strongest objection with a reason.
Explain precisely how the new information changes, or does not change, your decision.

I would move from a launch to a limited, non-decisive pilot. Faster processing matters, but the new evidence questions what our success label measures. Promotion may reflect access to mentoring as well as ability, and broad training coverage would not solve that alone. We should evaluate the work-sample evidence, examine regional errors and give applicants an independent assessment route. The model could assist administration while those issues are addressed, but it should not automatically reject people. I would reconsider wider use when the target and error costs are defensible.

Spoken exit

Your spoken exit

Give one question about data and one about the consequences of a model’s errors.

One possible response

Who is missing or misrepresented in the training examples? And who loses an opportunity when the model gets the decision wrong?

Your takeaway

One-minute takeaway

Choose one answer. Then explain your choice in one sentence aloud.

Can an Algorithm Be Fair?

Start talking

Clip A

Useful English

Clip B

Useful English

Mastery

The hiring committee

Spoken exit

Your takeaway

Lesson aims

Practice audio