A company rejects a job applicant.
A bank denies a loan.
An insurer flags a claim.
A landlord screens out a prospective tenant.
A hospital prioritizes one patient over another.
For most of modern history, someone could eventually be identified as the person who made the decision.
A manager.
A loan officer.
An underwriter.
A property owner.
A doctor.
Artificial intelligence is making that picture more complicated.
Increasingly, software can rank, recommend, predict, score, filter or flag people before a human being makes the final decision—and sometimes the human may do little more than accept what the system recommends.
When the decision is correct, the distinction may attract little attention.
When it is wrong, another question appears:
Who is responsible for a decision that came from a machine?
The Computer Usually Is Not the Legal Person
Artificial intelligence can perform tasks that look remarkably independent.
It can evaluate information.
Recognize patterns.
Generate recommendations.
Predict outcomes.
Rank applicants.
Flag suspicious behavior.
But an AI system is generally not a legal person that can simply assume responsibility for what it does.
The software was developed by someone.
Purchased by someone.
Configured by someone.
Given data by someone.
Deployed for a particular purpose by someone.
And its output is ultimately used by a person, business, agency or other organization.
That means the law usually has to look beyond the machine.
The important question is not merely:
What did the algorithm do?
It is:
Who used it, for what purpose, under what legal duty, and with what consequences?
A Company Cannot Necessarily Escape the Law by Automating the Decision
Imagine an employer is prohibited from discriminating against applicants based on a protected characteristic.
The employer begins using an automated hiring system.
The software evaluates applications and consistently disadvantages a protected group.
The employer cannot necessarily answer a discrimination claim simply by saying:
“We didn't discriminate. The computer made the decision.”
The decision-making process still belongs to the employer.
Federal employment laws continue to apply when employers use automated systems.
The Equal Employment Opportunity Commission has specifically addressed the application of federal anti-discrimination law to software, algorithms and artificial intelligence used in employment decisions.
The underlying principle is straightforward:
Automating a decision does not automatically remove the legal obligations attached to that decision.
Hiring Is Becoming One of the Most Visible Examples
Employers can receive hundreds or thousands of applications for a single position.
Automation offers an obvious solution.
Software can screen résumés.
Rank applicants.
Analyze assessments.
Compare qualifications.
Schedule interviews.
Some systems may evaluate recorded interviews or other applicant information.
This can make hiring much faster.
It can also create a new kind of opacity.
An applicant may receive a rejection without knowing whether a person ever meaningfully reviewed the application.
The employer may know that the software assigned a low score but not fully understand why.
The vendor may understand the system better but know almost nothing about the individual applicant.
Responsibility becomes distributed across several participants.
The Vendor and the Employer May Have Very Different Roles
Suppose Company A develops an automated hiring product.
Company B purchases it.
Company B decides which jobs will use it.
Company B supplies criteria.
The software ranks applicants.
Company B rejects everyone below a particular score.
If the system produces unlawful discrimination, who caused it?
Perhaps the design of the software contributed.
Perhaps Company B configured it poorly.
Perhaps the training data created a problem.
Perhaps the employer used the tool for a purpose for which it was never designed.
Perhaps several of those things happened simultaneously.
This is why AI liability is unlikely to fit comfortably into a universal rule that says either “the developer is responsible” or “the user is responsible.”
Responsibility can depend on what each participant actually did.
A Human Clicking “Approve” Does Not Necessarily Make the Decision Human
Organizations sometimes describe automated systems as merely advisory because a person technically makes the final decision.
But the amount of human involvement matters.
Imagine an algorithm evaluates 20,000 applications.
It rejects 19,000.
A manager reviews the remaining 1,000.
The company might accurately say that humans selected every person who was ultimately hired.
But the algorithm still determined which 19,000 applicants the humans never considered.
The automated system made an important decision even if it did not make the last one.
The same problem can occur elsewhere.
An algorithm flags an insurance claim.
A human investigator then scrutinizes it.
A risk model gives someone a low score.
A human loan officer relies on that score.
A fraud system freezes a transaction.
A human reviews the matter only after the customer complains.
Human involvement does not necessarily erase the effect of the automated decision.
Artificial Intelligence Can Reproduce Patterns Hidden in Historical Data
One of AI's greatest strengths is its ability to find patterns.
That can also create risk.
Imagine a company trains a hiring model using information about people who were historically successful at the company.
The system searches for characteristics associated with those workers.
That sounds sensible.
But what if the company's historical workforce reflects decades of unequal hiring?
The algorithm may discover patterns that correlate with past employment decisions without understanding why those patterns existed.
It does not need to contain an explicit instruction saying:
“Prefer this demographic group.”
It can potentially rely on other characteristics that correlate with the historical pattern.
The computer sees relationships in data.
It does not independently decide whether those relationships are fair, lawful or socially desirable.
Removing a Protected Characteristic Does Not Always Remove Its Influence
Suppose a lender removes race from a dataset.
Or an employer removes sex.
At first glance, that seems to eliminate the possibility that the system will consider that characteristic.
But other information can sometimes correlate with characteristics that were removed.
Location.
Employment history.
Education.
Purchasing patterns.
Language.
Organizations therefore cannot always assume that removing one field eliminates every pathway through which a model could produce a problematic disparity.
The legal analysis depends on the particular law and facts, but the technical lesson is broader.
A machine can find patterns humans did not deliberately program.
That is part of what makes machine learning useful.
It is also part of what makes oversight difficult.
Credit Decisions Already Have Rules About Explanations
Credit provides an especially useful example because automated decision-making did not begin with modern generative AI.
Lenders have used scoring and automated systems for decades.
Federal law nevertheless imposes requirements concerning adverse credit decisions.
Under the Equal Credit Opportunity Act and Regulation B, creditors generally must provide applicants with specific reasons for certain adverse actions or disclose the right to obtain those reasons.
The Consumer Financial Protection Bureau has emphasized that creditors do not escape those obligations merely because they use complex algorithms.
A creditor using sophisticated technology still has to comply with applicable requirements for explaining adverse action.
That creates an important principle for the AI era:
Complexity inside the machine does not necessarily eliminate an organization's obligation to explain what it did.
“The Algorithm Did It” Is Not Much of an Explanation
Imagine someone applies for a mortgage.
The application is denied.
The applicant asks why.
The bank responds:
“Our artificial-intelligence system determined that you did not qualify.”
That tells the applicant almost nothing.
Which information mattered?
Income?
Debt?
Credit history?
Something else?
If the law requires reasons, identifying the existence of an algorithm is not necessarily the same thing as identifying the reason for the decision.
This becomes harder as models become more complex.
Some automated systems can contain enormous numbers of interacting variables.
Their predictions can be highly accurate while remaining difficult to translate into an ordinary explanation.
That creates a tension between technological sophistication and legal accountability.
A Highly Accurate System Can Still Be Wrong About One Person
Suppose a fraud-detection model correctly identifies suspicious transactions 99 percent of the time.
That would be extraordinarily useful.
But the remaining errors involve real people.
Someone's legitimate payment gets blocked.
An account is frozen.
A claim is delayed.
A transaction is reported for additional review.
Aggregate accuracy does not erase the consequences of an individual mistake.
This distinction matters whenever organizations defend automated systems by pointing to overall performance.
The system can be excellent in general and still be wrong in a particular case.
The person affected by that particular error experiences the individual decision, not the statistical average.
Insurance Creates Similar Questions
Insurance depends heavily on prediction.
How likely is a driver to have an accident?
How much will a property cost to insure?
Does a claim appear unusual?
How much risk does a particular policy represent?
Algorithms are naturally suited to these questions because insurers possess large amounts of data.
AI can potentially make underwriting and claims processing faster and more precise.
But it also raises questions about what information is being used.
If an automated system affects premiums, coverage or claim handling, regulators and courts may eventually need to understand how the decision was made and whether the use of particular information complied with applicable law.
Again, the fact that a machine performed the analysis does not move the decision outside the legal system.
Housing Decisions Can Be Automated Before a Landlord Ever Sees an Applicant
Rental screening is another area where software can exert enormous influence.
A landlord may use a third-party screening company.
The system can evaluate credit information, rental history, records supplied by databases and other criteria.
The landlord may receive a recommendation rather than the complete underlying information.
That creates several potential points of failure.
The underlying record may belong to someone else.
A record may be outdated.
A database may contain an error.
A scoring model may interpret information in a way the applicant cannot see.
The landlord may rely heavily on the recommendation.
From the applicant's perspective, the result is simple:
The application is denied.
Behind that simple result may be a chain involving several companies and automated processes.
Bad Data Can Produce a Perfectly Functioning Wrong Answer
Not every AI failure is actually an algorithm failure.
Sometimes the algorithm performs exactly as designed.
The information it receives is wrong.
Imagine a system evaluating a person based on a criminal record belonging to someone with a similar name.
The model accurately processes the record.
The score is calculated correctly.
The recommendation follows the programmed rules.
And the result is still wrong because the starting information concerned somebody else.
This is an important distinction.
When an automated decision causes harm, investigators may need to determine whether the problem came from:
The data.
The model.
The way the model was configured.
The way the output was interpreted.
Or the final action taken by the organization.
Calling all of these an “AI error” can hide what actually happened.
The Machine May Know Correlation Without Knowing Cause
Suppose an algorithm discovers that people with characteristic X are more likely to default on a loan.
That may be statistically true in the data.
But why?
Is X actually connected to creditworthiness?
Does it merely correlate with another factor?
Is the historical data distorted?
Does the relationship continue to exist?
Could using X violate the law?
Machine-learning systems are extraordinarily good at discovering correlations.
Legal decision-making often requires something more.
It may require asking whether the factor is permissible to use at all.
The algorithm cannot resolve that legal question simply because the mathematical relationship is strong.
Government Use Raises Constitutional Questions Too
The stakes become even greater when government agencies use automated systems.
An algorithm might assist with fraud detection, benefits administration, law enforcement or other public functions.
Government action can implicate constitutional protections and administrative-law requirements in ways private decisions may not.
If a government system contributes to depriving someone of property, benefits or liberty, questions of notice and an opportunity to challenge the decision can become particularly important.
A person cannot meaningfully contest an error if nobody can explain how the decision occurred.
That makes transparency more than a technical preference in some government contexts.
It can become part of procedural fairness.
The Developer May Never Meet the Person Harmed
Traditional negligence often involves people who can easily identify one another.
A driver hits another vehicle.
A doctor treats a patient.
A contractor builds a house.
AI can create much longer chains.
A developer writes software.
A technology company trains a model.
A vendor incorporates it into a product.
A business buys the product.
An employee configures it.
The system evaluates a consumer.
The consumer suffers the consequence.
The developer at the beginning of the chain may never know the consumer exists.
Determining responsibility requires tracing what happened through that chain.
Contracts Can Allocate Responsibility Without Necessarily Ending It
Businesses buying AI systems often enter detailed contracts with vendors.
Those agreements may address warranties, limitations of liability, indemnification, data responsibilities and permitted uses.
Such provisions can determine who ultimately bears financial responsibility between the companies.
But a contract between two businesses does not necessarily eliminate rights that another person may have under applicable law.
Imagine a vendor agrees to reimburse an employer for certain legal claims arising from its software.
That contractual arrangement may matter enormously between the vendor and employer.
It does not necessarily determine whether the rejected job applicant has a valid claim in the first place.
There can therefore be two layers of responsibility:
Responsibility to the person harmed.
And responsibility between the businesses involved.
AI Makes Documentation More Important
When a human makes a consequential decision, that person can sometimes explain the reasoning later.
An automated process may evaluate thousands of cases every hour.
If the organization does not preserve meaningful records, reconstructing one particular decision months later can become difficult.
Which version of the model was operating?
What data did it receive?
What score did it produce?
What threshold was used?
Was the result overridden?
Did a human review it?
Was the model updated afterward?
Those details can matter when a decision is challenged.
Without records, an organization may know what happened—the application was denied—without being able to reconstruct why it happened.
AI Can Also Make Decisions More Consistent
Automation is not inherently less fair than human judgment.
Human decision-makers have biases.
They become tired.
They can treat similar cases differently.
They can overlook information.
They can make mathematical mistakes.
A properly designed automated system can apply criteria consistently across enormous numbers of cases.
It can detect patterns humans would miss.
It can flag errors.
It can reduce arbitrary variation.
That is part of the reason organizations adopt these systems.
The legal challenge is not to assume that automated decisions are inherently bad.
It is to determine how existing protections apply when an automated system produces a decision that would be unlawful if a human made it.
The Question Is Shifting From “Did AI Decide?” to “How Was AI Used?”
Artificial intelligence is becoming integrated into ordinary software.
Soon, asking whether a decision “used AI” may be almost meaningless.
A human may use software containing dozens of automated components.
One system retrieves information.
Another predicts risk.
Another ranks possibilities.
Another generates an explanation.
A person makes the final selection.
Which part made the decision?
The better question may be how each system influenced the result.
That approach focuses on what actually happened rather than whether a product carries an AI label.
Responsibility Does Not Disappear Inside the Machine
Organizations have always used tools to make decisions.
Calculators.
Credit scores.
Databases.
Background checks.
Statistical models.
Search systems.
Artificial intelligence is a much more powerful continuation of that history.
What makes the current moment different is the degree of autonomy these systems can appear to possess.
The software can produce an answer so quickly and confidently that the answer begins to feel like an independent judgment.
But the legal system still operates in a world of people and organizations.
Someone chose to deploy the system.
Someone decided what role its output would play.
Someone acted on the result.
And when that result affects employment, housing, credit, insurance, government benefits or other important interests, the existence of an algorithm does not make the consequences disappear.
The most important question may therefore be less futuristic than it sounds.
When artificial intelligence makes a consequential decision, the law still has to find the human institutions standing behind the machine.