Speaker: 00:00:00

This week's article is Consensus

Without Consequence, the

2

: 00:00:03

Collapse of AI Accountability.

3

: 00:00:07

There is a sentence that has been

true for nearly a decade, and the

4

: 00:00:10

fact that it remains true is, by now,

something close to an indictment.

5

: 00:00:17

Everyone agrees that artificial

intelligence should be fair,

6

: 00:00:20

transparent, and accountable.

7

: 00:00:24

You could have read that sentence in 2018.

8

: 00:00:27

You could have read it

in: 2020

9

: 00:00:32

2024.

10

: 00:00:34

The words have not changed.

11

: 00:00:36

The situation they

describe has not changed.

12

: 00:00:41

What has changed is our ability

to pretend that agreeing on the

13

: 00:00:44

words was ever the difficult part.

14

: 00:00:47

A landmark review by Anna Jobin, Marcello

Ienca, and Effy Vayena, examining more

15

: 00:00:52

than two hundred AI ethics guidelines

and governance documents from around

16

: 00:00:56

the world, found that transparency

appeared in 86 per cent of them.

17

: 00:01:01

Justice and fairness in 81 per cent.

18

: 00:01:04

Non-maleficence in 71 per cent.

19

: 00:01:07

The world, it turns out, has been

extraordinarily good at articulating

20

: 00:01:12

what responsible AI ought to involve.

21

: 00:01:16

The world has been catastrophically

bad at enforcing it.

22

: 00:01:20

That gap — between articulation

and enforcement — is not

23

: 00:01:25

an abstract policy debate.

24

: 00:01:27

It is the difference between a hiring

algorithm that discriminates against

25

: 00:01:31

older workers and one that does not.

26

: 00:01:35

It is the difference between a

facial recognition system that

27

: 00:01:38

operates with impunity and one

that faces genuine consequences.

28

: 00:01:44

It is the difference between an

ethics board that exists to absorb

29

: 00:01:47

criticism and one that has the

power to halt a product launch.

30

: 00:01:53

The question that actually matters now is

deceptively simple: what does meaningful

31

: 00:01:59

accountability look like in practice?

32

: 00:02:02

And when enforcement fails to

materialise in time, who bears the cost?

33

: 00:02:08

The proliferation of ethics guidelines

over the past decade represents one

34

: 00:02:12

of the most remarkable exercises

in global norm-setting since the

35

: 00:02:16

Universal Declaration of Human Rights.

36

: 00:02:19

Governments, corporations,

academic institutions, and

37

: 00:02:23

civil society organisations have

produced hundreds of frameworks.

38

: 00:02:28

The World Economic Forum has described

the challenge as turning ethical

39

: 00:02:31

principles into tangible practices.

40

: 00:02:35

The International Labour Organization

has reviewed global guidelines

41

: 00:02:39

specifically for AI in the workplace,

finding consistent themes around

42

: 00:02:43

worker protection and human oversight.

43

: 00:02:47

The apparent consensus is real.

44

: 00:02:50

And it masks a deeper dysfunction.

45

: 00:02:53

As research published in the journal

Patterns noted, while the most advocated

46

: 00:02:57

ethical principles show significant

convergence, there remains, and this is

47

: 00:03:02

the crucial detail, substantive divergence

in how those principles are interpreted,

48

: 00:03:07

why they are deemed important, what

domains and actors they apply to, and

49

: 00:03:12

how they should actually be implemented.

50

: 00:03:15

Everyone agrees on the words.

51

: 00:03:18

Nobody agrees on what the

words mean in practice.

52

: 00:03:22

This is the principles paradox.

53

: 00:03:24

The more guidelines that exist, the easier

it becomes for organisations to claim

54

: 00:03:28

alignment with ethical AI whilst doing

very little to change their behaviour.

55

: 00:03:33

The phenomenon has a name: ethics washing.

56

: 00:03:37

And it has become, in 2025 and

: 2026

57

: 00:03:42

the corporate AI landscape.

58

: 00:03:44

When a company publishes a set

of ethics principles, appoints

59

: 00:03:47

a chief ethics officer, and then

deploys systems that systematically

60

: 00:03:51

discriminate, the principles

themselves become a form of camouflage.

61

: 00:03:55

A shield against criticism rather

than a genuine constraint on conduct.

62

: 00:04:00

The most notorious illustration of

played out at Google in late: 2020

63

: 00:04:04

Timnit Gebru, co-lead of Google's Ethical

AI team, was fired after the company

64

: 00:04:10

demanded she retract a research paper

examining the environmental costs and

65

: 00:04:15

bias risks of large language models.

66

: 00:04:18

Three months later, Margaret Mitchell,

the team's founder, was also terminated.

67

: 00:04:23

Roughly 2,700 Google employees and more

than 4,300 academics and civil society

68

: 00:04:30

supporters signed letters of condemnation.

69

: 00:04:34

The paper that triggered the dispute

— "On the Dangers of Stochastic Parrots:

70

: 00:04:38

Can Language Models Be Too Big?"

71

: 00:04:41

— was subsequently presented at a major

academic conference and has since become

72

: 00:04:46

one of the most cited works in the field.

73

: 00:04:49

The episode demonstrated something

that has only become clearer with

74

: 00:04:52

time: internal ethics teams cannot

function as accountability mechanisms

75

: 00:04:57

when they exist at the pleasure of the

organisations they are meant to constrain.

76

: 00:05:02

The fox does not appoint

its own gamekeeper.

77

: 00:05:05

The numbers that have emerged from

industry surveys are stark in a

78

: 00:05:09

different register — not dramatic,

but relentlessly cumulative.

79

: 00:05:14

According to ISACA's 2025 global survey

of more than 3,200 business and IT

80

: 00:05:21

professionals, nearly three out of

four European IT and cybersecurity

81

: 00:05:26

professionals reported that staff

were already using generative AI at

82

: 00:05:30

work, a figure that had risen ten

percentage points in a single year.

83

: 00:05:36

Yet only 31 per cent of

organisations had a formal,

84

: 00:05:39

comprehensive AI policy in place.

85

: 00:05:42

Sixty-three per cent were extremely

or very concerned that generative

86

: 00:05:47

AI could be weaponised against their

87

: 00:05:50

organisations.

88

: 00:05:50

Eighteen per cent had invested

in tools to address that concern.

89

: 00:05:54

A separate analysis found that 57 per

cent of organisations acknowledged

90

: 00:05:58

that AI was advancing more

quickly than they could secure it.

91

: 00:06:02

The phrase "governance gap" has

become a staple of policy discourse.

92

: 00:06:06

It undersells the scale of the problem.

93

: 00:06:09

This is not a gap.

94

: 00:06:11

It is a chasm.

95

: 00:06:12

The Partnership on AI identified

six governance priorities for: 2026

96

: 00:06:18

responsible adoption of agentic AI

systems, improved documentation and

97

: 00:06:23

transparency standards, governance

convergence across jurisdictions,

98

: 00:06:29

protections for authentic human

voice in an era of synthetic content.

99

: 00:06:35

The priorities are sensible.

100

: 00:06:37

They are also an implicit admission

that none of these foundations are yet

101

: 00:06:40

in place, despite years of discussion.

102

: 00:06:44

Meanwhile, agentic AI systems — which

take autonomous actions in the real

103

: 00:06:49

world rather than simply generating

text — introduce what the Partnership

104

: 00:06:54

describes as non-reversibility of

actions, open-ended decision-making

105

: 00:07:00

pathways, and privacy vulnerabilities

from expanded data access.

106

: 00:07:06

These are not theoretical risks.

107

: 00:07:08

They are features of systems already

deployed in customer service, software

108

: 00:07:12

development, and financial trading.

109

: 00:07:16

The governance frameworks meant

to constrain them are, in many

110

: 00:07:19

cases, still being drafted.

111

: 00:07:22

The European Union's AI Act

represents the most ambitious

112

: 00:07:25

attempt to date to translate ethical

principles into enforceable law.

113

: 00:07:30

It entered into force in August

: 2024

114

: 00:07:35

timeline extending through 2027.

115

: 00:07:39

Prohibitions on the most dangerous AI

tions took effect in February: 2025

116

: 00:07:45

Full enforcement of requirements for

high-risk systems — with fines reaching

117

: 00:07:49

up to 35 million euros or seven per

cent of global annual turnover — does

118

: 00:07:54

not arrive until August 2026.

119

: 00:07:58

This is, by any measure, a

significant regulatory achievement.

120

: 00:08:03

But the Act was first

proposed in April: 2021

121

: 00:08:07

When the European Commission drafted

that proposal, ChatGPT did not exist.

122

: 00:08:13

Nor did the current generation of

autonomous agents, multimodal models,

123

: 00:08:17

or AI-powered code generation tools.

124

: 00:08:21

The regulation is, by design,

chasing a target that moved while

125

: 00:08:25

lawmakers were still aiming.

126

: 00:08:27

The United States presents a

different set of challenges entirely.

127

: 00:08:31

Rather than pursuing comprehensive

legislation, it has relied on a

128

: 00:08:34

decentralised approach combining

agency-specific enforcement, voluntary

129

: 00:08:39

frameworks, and sector-level regulation.

130

: 00:08:43

Then, in December 2025, President Trump

signed an executive order seeking what the

131

: 00:08:48

administration described as a minimally

burdensome national policy framework.

132

: 00:08:54

The order directed the Attorney

General to establish an AI Litigation

133

: 00:08:57

Task Force to challenge state AI

laws deemed inconsistent with federal

134

: 00:09:02

policy, and instructed the Secretary

of Commerce to identify state

135

: 00:09:06

legislation considered "onerous."

136

: 00:09:09

It even tied federal broadband

infrastructure funding to compliance

137

: 00:09:13

with those determinations.

138

: 00:09:15

The order was, in effect, an

attempt to pre-empt a patchwork of

139

: 00:09:18

state-level regulations that had

been emerging with genuine ambition.

140

: 00:09:23

Colorado's legislation, effective February

: 2026

141

: 00:09:29

of high-risk AI systems to use reasonable

care to protect consumers from algorithmic

142

: 00:09:35

discrimination, implement risk management

policies, and conduct impact assessments.

143

: 00:09:41

New York City had already established

bias audit requirements for

144

: 00:09:45

automated employment decision tools.

145

: 00:09:47

More than a hundred state AI

laws were enacted across the

146

: 00:09:50

United States in 2025 alone.

147

: 00:09:53

Governors in California, Colorado,

and New York indicated they would

148

: 00:09:57

enforce their statutes regardless.

149

: 00:10:00

Legal scholars noted the constitutional

questions were substantial.

150

: 00:10:05

The result is a governance landscape that

is not merely fragmented but actively

151

: 00:10:09

contested, with federal and state

authorities pulling in opposing directions

152

: 00:10:14

whilst companies navigate overlapping

and sometimes contradictory obligations.

153

: 00:10:19

When enforcement mechanisms fail to

materialise in time, the costs do

154

: 00:10:24

not distribute themselves evenly.

155

: 00:10:27

They concentrate, with brutal

predictability, on those with

156

: 00:10:32

the least power to resist.

157

: 00:10:34

In employment, five individuals

over the age of forty applied

158

: 00:10:38

for hundreds of positions through

Workday's automated hiring platform

159

: 00:10:43

and received almost no interviews.

160

: 00:10:46

They alleged that Workday's

AI recommendation system

161

: 00:10:49

discriminated on the basis of age.

162

: 00:10:52

In 2024, a court allowed the disparate

impact claim to proceed, holding that

163

: 00:10:58

Workday bore liability as an agent

of the employers using its product.

164

: 00:11:03

In housing, plaintiffs demonstrated

that the SafeRent tenant screening

165

: 00:11:08

algorithm produced discriminatory

outcomes for Black and Hispanic

166

: 00:11:12

applicants, and the company settled for

e than two million dollars in: 2024

167

: 00:11:18

In biometric surveillance, Clearview

AI scraped billions of photographs

168

: 00:11:23

from social media without consent,

sold facial recognition services to law

169

: 00:11:27

enforcement worldwide, was fined 30.5

170

: 00:11:31

million euros by the Dutch data

protection authority, and then settled a

171

: 00:11:35

US class action for approximately 51.75

172

: 00:11:39

million dollars — structured,

extraordinarily, as a 23 per

173

: 00:11:44

cent equity stake in the company

itself, because Clearview had

174

: 00:11:48

insufficient assets to pay in cash.

175

: 00:11:51

A bipartisan group of state attorneys

general filed formal objections

176

: 00:11:55

to the settlement's adequacy.

177

: 00:11:57

These cases share a common structure.

178

: 00:12:00

Harm occurs.

179

: 00:12:01

Years pass.

180

: 00:12:03

Legal proceedings unfold.

181

: 00:12:05

Settlements are reached or fines imposed.

182

: 00:12:09

But the systems that caused the harm

often continue operating throughout

183

: 00:12:13

the entire adjudication process, and

the individuals affected rarely receive

184

: 00:12:18

compensation proportional to their injury.

185

: 00:12:21

The enforcement mechanisms

exist, technically.

186

: 00:12:25

They simply do not work fast

enough to prevent the damage

187

: 00:12:27

they are meant to address.

188

: 00:12:30

A study from the University of Washington

provided evidence of the scale of

189

: 00:12:33

algorithmic bias in employment contexts.

190

: 00:12:36

Researchers presented three AI models

with job applications identical in every

191

: 00:12:41

respect except the name of the applicant.

192

: 00:12:44

The models preferred resumes with

white-associated names in 85 per cent

193

: 00:12:48

of cases and those with Black-associated

names only 9 per cent of the time.

194

: 00:12:54

A separate study published in June 2025,

led by researchers at Cedars-Sinai,

195

: 00:13:00

found that leading large language models

generated less effective treatment

196

: 00:13:04

recommendations when a patient's race

was identified as African American.

197

: 00:13:09

These are not edge cases.

198

: 00:13:11

They are documented patterns

of discriminatory behaviour

199

: 00:13:14

embedded in systems that millions

of people interact with daily.

200

: 00:13:19

And they persist not because the ethical

principles are inadequate, but because

201

: 00:13:23

the mechanisms for enforcing those

principles remain woefully underdeveloped.

202

: 00:13:28

Algorithmic auditing is often proposed

as the solution: independent third

203

: 00:13:32

parties evaluating AI systems for

bias and compliance, much as financial

204

: 00:13:37

auditors examine corporate accounts.

205

: 00:13:40

New York City requires annual bias audits

for automated employment decision tools.

206

: 00:13:46

Colorado mandates impact

assessments for high-risk systems.

207

: 00:13:51

The EU AI Act requires conformity

assessments for high-risk applications.

208

: 00:13:57

The AI Now Institute has mounted a

detailed critique of this approach,

209

: 00:14:02

arguing that technical evaluations

narrowly position bias as a flaw that

210

: 00:14:06

can be fixed and eliminated, when in fact

algorithmic harms are often structural,

211

: 00:14:13

reflecting the social contexts in which

systems are designed and deployed.

212

: 00:14:18

Audits, the institute contends,

risk entrenching power within the

213

: 00:14:22

technology industry whilst taking focus

away from more structural responses.

214

: 00:14:27

The critique has substance.

215

: 00:14:29

There are no universally

accepted standards for what

216

: 00:14:32

constitutes a passing score.

217

: 00:14:35

Audit costs range from five

thousand to fifty thousand pounds

218

: 00:14:39

depending on system complexity.

219

: 00:14:42

Audits evaluate systems at a single

point in time, but AI models drift

220

: 00:14:47

as they encounter new data, meaning a

system that passes today may produce

221

: 00:14:52

discriminatory outcomes next month.

222

: 00:14:56

And audits place the primary

burden of accountability on

223

: 00:14:59

those with the fewest resources.

224

: 00:15:02

The information asymmetry is

profound and, under current

225

: 00:15:05

frameworks, largely unaddressed.

226

: 00:15:08

The geopolitical dimension

complicates everything further.

227

: 00:15:12

At the AI Action Summit

in Paris in February: 2025

228

: 00:15:17

58 nations signed a joint

declaration on inclusive and

229

: 00:15:20

sustainable artificial intelligence.

230

: 00:15:23

The United States and the

United Kingdom refused to sign.

231

: 00:15:27

Anthropic's chief executive described

the summit as a missed opportunity

232

: 00:15:32

for addressing AI safety, reflecting a

broader frustration that international

233

: 00:15:37

forums produce declarations

rather than binding commitments.

234

: 00:15:42

Meanwhile, researchers have documented

how AI development mirrors historical

235

: 00:15:46

patterns of colonial resource extraction.

236

: 00:15:49

Control over data infrastructures,

computational resources, and

237

: 00:15:53

algorithmic systems remains

concentrated in a small number of

238

: 00:15:57

wealthy nations and corporations.

239

: 00:16:00

Environmental costs fall

disproportionately on regions where data

240

: 00:16:04

centres proliferate because electricity

and land are cheap, exporting the

241

: 00:16:09

benefits of artificial intelligence

whilst localising its burdens.

242

: 00:16:14

When 98 per cent of AI research originates

from wealthy institutions, the resulting

243

: 00:16:19

systems embed assumptions that may

be irrelevant or damaging elsewhere.

244

: 00:16:24

The world is not building

a shared technology.

245

: 00:16:27

It is building one that reflects

the interests of those who built it.

246

: 00:16:31

What would a more effective

system actually require?

247

: 00:16:35

The evidence points to several structural

necessities that go beyond the familiar

248

: 00:16:39

call for more principles or better audits.

249

: 00:16:43

Accountability must be anticipatory

rather than reactive — the current

250

: 00:16:47

model waits for harm to occur, then

attempts to assign responsibility through

251

: 00:16:52

litigation or regulatory action long

after damage has accumulated across

252

: 00:16:56

thousands of individual decisions.

253

: 00:17:00

Enforcement must be resourced

proportionally to the scale of deployment.

254

: 00:17:04

The finding that only 31 per cent

of organisations have comprehensive

255

: 00:17:07

AI policies is not simply a failure

of corporate governance; it reflects

256

: 00:17:13

a reality in which the institutions

responsible for oversight lack the

257

: 00:17:17

funding, technical expertise, and legal

authority to match the pace of industry.

258

: 00:17:23

Transparency must extend beyond model

documentation to encompass the full chain

259

: 00:17:28

of development and deployment, enabling

affected communities — not just regulators

260

: 00:17:32

and auditors — to understand how decisions

are made and what recourse is available.

261

: 00:17:39

The costs of non-compliance

must be sufficiently high to

262

: 00:17:41

alter corporate behaviour.

263

: 00:17:44

And governance frameworks must be

designed for iteration, not permanence,

264

: 00:17:48

because a five-year legislative cycle

is simply incompatible with a technology

265

: 00:17:52

that transforms every six months.

266

: 00:17:55

None of these requirements are novel.

267

: 00:17:57

Researchers, civil society

organisations, and some regulators have

268

: 00:18:01

been advocating for them for years.

269

: 00:18:04

The obstacle is not a lack of ideas.

270

: 00:18:08

It is a lack of political will,

complicated by the enormous economic

271

: 00:18:12

interests that benefit from the

current arrangement — in which

272

: 00:18:16

deployment runs ahead of governance,

and the costs of failure are borne by

273

: 00:18:20

those least equipped to absorb them.

274

: 00:18:23

The financial scale of what

has been allowed to occur

275

: 00:18:25

is staggering in aggregate.

276

: 00:18:28

Individual settlements and fines

may appear substantial in isolation.

277

: 00:18:32

Set against the revenues of the companies

deploying these systems, and against

278

: 00:18:36

the cumulative harm inflicted across

millions of affected individuals, they

279

: 00:18:41

represent a cost of doing business

rather than a meaningful deterrent.

280

: 00:18:46

The economics of non-compliance

remain, for the moment, firmly

281

: 00:18:50

in favour of deploying first

and accounting for it later.

282

: 00:18:53

The field of AI ethics has accomplished

something genuinely remarkable in building

283

: 00:18:58

global consensus around core values.

284

: 00:19:02

That achievement should not be dismissed.

285

: 00:19:05

But consensus without enforcement

is aspiration without consequence.

286

: 00:19:10

And aspiration without consequence

is just another way of saying

287

: 00:19:13

that nobody is responsible.