Sign In

Communications of the ACM

Last byte

Agent Algorithm

noir figure in hat at night, illustration

Credit: Lukeruk / Shutterstock

Agent Saskia Lipcott grimaced at her reflection in the mirror glass, wondering again if the look she'd gone for had too much of Dana Scully from "The X-Files." "I don't get it," she said. "Why call the computer Duppin? It sounds like a kid's game."

"Not Duppin," said Special Agent Dinah Carter. "That's Doo-pan. It's from Edgar Allan Poe. He wrote the first detective stories, about a Frenchman called Auguste Dupin."

"Okay?" Lipcott sounded uncertain.

"And it's not the name of the computer; it's the machine-learning software. Didn't they give you the AI guide?"

Lipcott, who had grown up on a farm and had a very different idea of what AI stood for, raised her eyebrows. "Nope."

Carter opened her desk drawer, took out a thin book, and threw it across. The AI Primer—Artificial Intelligence from DeepMind to Dupin. "Let's move," she said to Lipcott. "We need to interview the suspect. It's a formality. Dupin tells us he's guilty, but it has to be done."

Lipcott followed Carter into the interview room. It was far less impressive than the ones she had seen in TV dramas—just an ordinary, bare office. A middle-aged man with thinning brown hair, wearing a sweater and jeans, was seated at a metal-topped table.

Carter drew up one of the two chairs on their side of the table and indicated Lipcott should take the other.

"Mr. Lamb, I am Special Agent Carter and this is Agent Lipcott. The purpose of this meeting is to put before you the evidence provided that will be used in your trial. You have refused legal representation. Any comments you make will be recorded and added to the evidence base. Do you understand?"

Lamb nodded.

Carter touched her tablet screen. "We have video placing you at the scene of an armed robbery last Monday at a jewelry store on Connecticut Avenue Northwest, between Desales and L streets. You are clearly seen committing the crime. This evidence is backed up by location data from your cellphone, showing that you travelled from your apartment to the store, arriving at the exact time the crime was committed. DNA evidence from the shotgun makes it clear that you were involved.

"Do you have anything to say?"

"Only what I've said all along," said Lamb. "I was at home all evening. I never left the apartment. I did not commit a crime. I was alone, but I spoke several times to my smart speaker."

Carter countered. "Though the system has no record of this. Do you have anything else to add?"

Lamb shrugged. "What's the point? It's a stitch-up."

"Thank you for your cooperation," said Carter. She nodded to the door. "Okay Lipcott, I'll see you outside. I just need to authorize the suspect's removal."

"That's it?" said Lipcott a few minutes later, as Carter emerged into the corridor. Lipcott had been flicking through the AI primer and now pointed at a paragraph. "Have you read this about artificial intelligence learning to play games?"

"Sure, I've read it," said Carter. "They're real good at games. That's why Dupin's perfect; detection is a kind of game. Isn't that why you became an investigator?"

"I guess," said Lipcott. "But listen to this. 'The remarkable success of AI systems is even more impressive because they were never given the rules of the game. Initially, their attempts were random failures, but over time, the machine-learning algorithm picked up strategies that were successful, often discovering loopholes in the game software and so managing to win in a way that the game designers never envisaged.'

"Isn't that worrying?" Lipcott pressed.

"Why?" asked Carter. "It shows how clever the software is. It doesn't need to know the rules: it plays vast numbers of times and sees what improves its score."

"That's great for playing Breakout," said Lipcott, unconvinced by Carter's defense. "But what about Dupin? How did it learn to be so successful?"

"The same way. It was fed vast numbers of case details and worked out the rules for itself. You'll see when you read the book properly that in the early days of AI they tried to list all the rules for the computer, but it never worked well enough for a complex task like detection. Machine learning totally beats anything we can teach the system."

"Okay, sure. But think of that line…" Lipcott searched the page in front of her. "'Managing to win in a way the game designers never envisaged.' If you translate that into what Dupin does, isn't that saying, 'managing to convict someone in a way that the law is not intended to work?' Could Dupin break the rules to win?"

"Not at all," said Carter. "Look, in the old days we relied on witness statements, even though they were useless. Have you heard of the 1901 Stern experiment?"

Lipcott shook her head.

"This German professor faked a murder in his lecture room and then asked the students, who thought the crime was real, to write accounts of what they saw. They named eight different people as the murderer. We rely on solid evidence. Of course Dupin considers witness statements and monitors all conversations in this building to include interviews and our thinking on the case, but it's Dupin's ability to access substantial evidence that makes it so effective."

"I understand that. But this is an artificial intelligence that made up its own rules. What's to say that the video isn't a deepfake, or the location data corrupted? Do we really know what Dupin is capable of?"

"Let's get a coffee." Carter headed down the corridor. "It's hard to accept at first—it feels like we're not real investigators anymore. But at least we've got jobs. Think of everyone replaced by machine-learning systems. The fact is, Dupin does the spade work, and we can concentrate on what humans do best. Interpreting outcomes and interacting with people."

"I get it," said Lipcott. "But how can we trust an algorithm that doesn't know the rules?"

How can we trust an algorithm that doesn't know the rules?

"You'll get used to it," said Carter. A noisy fanfare came from her pocket. "Sorry, I should change that alert. It's the Dupin app. Looks like we've got another case."

She pulled her phone out as she walked and flicked up the details. "This is what I mean. There was a homicide just five minutes ago in E street, and Dupin has already cracked it. Have you ever been there?"

Lipcott shook her head.

"It's where the DC medical examiner is based. It's our very own Rue Morgue."

"I'm sorry?"

"The Rue Morgue in Paris. That's where Poe's Dupin solved his first crime." Carter flicked at her phone again. "The details are coming through."

For a moment, Carter seemed to slip on the polished floor of the old FBI building. The Special Agent stopped dead and put her hand on the wall to steady herself. In a flashing red rectangle, the app was presenting her with a case summary:

Suspect: Saskia Lipcott.

Crime: Homicide (first degree).

Probability of conviction: 99.99%.

"What is it?" asked Lipcott.

Carter shook her head and stared at the combined camera and microphone that surveyed the corridor. Lipcott's words seemed to float in front of her eyes. What she did next could determine not just Lipcott's future, but her own.

She walked on to the corner, to a dead spot between cameras, took a deep breath, and mouthed, "Don't ask questions. Get ready to run."

Back to Top


Brian Clegg ( is a science writer based in the U.K. His most recent books are What Do You Think You Are?, exploring the science of what makes you you, and Quantum Computing, offering background to this new computing paradigm.

©2021 ACM  0001-0782/21/8

Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and full citation on the first page. Copyright for components of this work owned by others than ACM must be honored. Abstracting with credit is permitted. To copy otherwise, to republish, to post on servers, or to redistribute to lists, requires prior specific permission and/or fee. Request permission to publish from or fax (212) 869-0481.

The Digital Library is published by the Association for Computing Machinery. Copyright © 2021 ACM, Inc.


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account
Article Contents: