The global agreement on AI ethics (fairness, transparency, accountability) has not translated into enforcement, creating a widening gap between principles and practice.

Reviews of hundreds of guidelines show strong convergence on stated values, but major divergence on interpretation and implementation, enabling “ethics washing,” illustrated by Google’s 2020 firing of Timnit Gebru and later Margaret Mitchell.

Industry adoption of generative AI is rapid while governance lags, especially as agentic systems spread. Regulatory responses are uneven: the EU AI Act phases enforcement through 2027, while the US is fragmented and contested between federal policy and state laws like Colorado and NYC rules. Real-world harms persist in hiring, housing, and biometric surveillance (Workday, SafeRent, Clearview), with slow legal remedies and documented bias in studies.

Audits are costly, time-limited, and structurally insufficient, and there is critical need for anticipatory, well-resourced, iterative governance with meaningful penalties and broader transparency.

This week's article is Consensus

Without Consequence, the

2

Collapse of AI Accountability.

3

There is a sentence that has been

true for nearly a decade, and the

4

fact that it remains true is, by now,

something close to an indictment.

5

Everyone agrees that artificial

intelligence should be fair,

6

transparent, and accountable.

7

You could have read that sentence in 2018.

8

You could have read it

9

2024.

10

The words have not changed.

11

The situation they

describe has not changed.

12

What has changed is our ability

to pretend that agreeing on the

13

words was ever the difficult part.

14

A landmark review by Anna Jobin, Marcello

Ienca, and Effy Vayena, examining more

15

than two hundred AI ethics guidelines

and governance documents from around

16

the world, found that transparency

appeared in 86 per cent of them.

17

Justice and fairness in 81 per cent.

18

Non-maleficence in 71 per cent.

19

The world, it turns out, has been

extraordinarily good at articulating

20

what responsible AI ought to involve.

21

The world has been catastrophically

bad at enforcing it.

22

That gap — between articulation

and enforcement — is not

23

an abstract policy debate.

24

It is the difference between a hiring

algorithm that discriminates against

25

older workers and one that does not.

26

It is the difference between a

facial recognition system that

27

operates with impunity and one

that faces genuine consequences.

28

It is the difference between an

ethics board that exists to absorb

29

criticism and one that has the

power to halt a product launch.

30

The question that actually matters now is

deceptively simple: what does meaningful

31

accountability look like in practice?

32

And when enforcement fails to

materialise in time, who bears the cost?

33

The proliferation of ethics guidelines

over the past decade represents one

34

of the most remarkable exercises

in global norm-setting since the

35

Universal Declaration of Human Rights.

36

Governments, corporations,

academic institutions, and

37

civil society organisations have

produced hundreds of frameworks.

38

The World Economic Forum has described

the challenge as turning ethical

39

principles into tangible practices.

40

The International Labour Organization

has reviewed global guidelines

41

specifically for AI in the workplace,

finding consistent themes around

42

worker protection and human oversight.

43

The apparent consensus is real.

44

And it masks a deeper dysfunction.

45

As research published in the journal

Patterns noted, while the most advocated

46

ethical principles show significant

convergence, there remains, and this is

47

the crucial detail, substantive divergence

in how those principles are interpreted,

48

why they are deemed important, what

domains and actors they apply to, and

49

how they should actually be implemented.

50

Everyone agrees on the words.

51

Nobody agrees on what the

words mean in practice.

52

This is the principles paradox.

53

The more guidelines that exist, the easier

it becomes for organisations to claim

54

alignment with ethical AI whilst doing

very little to change their behaviour.

55

The phenomenon has a name: ethics washing.

56

And it has become, in 2025 and

57

the corporate AI landscape.

58

When a company publishes a set

of ethics principles, appoints

59

a chief ethics officer, and then

deploys systems that systematically

60

discriminate, the principles

themselves become a form of camouflage.

61

A shield against criticism rather

than a genuine constraint on conduct.

62

The most notorious illustration of

63

Timnit Gebru, co-lead of Google's Ethical

AI team, was fired after the company

64

demanded she retract a research paper

examining the environmental costs and

65

bias risks of large language models.

66

Three months later, Margaret Mitchell,

the team's founder, was also terminated.

67

Roughly 2,700 Google employees and more

than 4,300 academics and civil society

68

supporters signed letters of condemnation.

69

The paper that triggered the dispute

— "On the Dangers of Stochastic Parrots:

70

Can Language Models Be Too Big?"

71

— was subsequently presented at a major

academic conference and has since become

72

one of the most cited works in the field.

73

The episode demonstrated something

that has only become clearer with

74

time: internal ethics teams cannot

function as accountability mechanisms

75

when they exist at the pleasure of the

organisations they are meant to constrain.

76

The fox does not appoint

its own gamekeeper.

77

The numbers that have emerged from

industry surveys are stark in a

78

different register — not dramatic,

but relentlessly cumulative.

79

According to ISACA's 2025 global survey

of more than 3,200 business and IT

80

professionals, nearly three out of

four European IT and cybersecurity

81

professionals reported that staff

were already using generative AI at

82

work, a figure that had risen ten

percentage points in a single year.

83

Yet only 31 per cent of

organisations had a formal,

84

comprehensive AI policy in place.

85

Sixty-three per cent were extremely

or very concerned that generative

86

AI could be weaponised against their

87

organisations.

88

Eighteen per cent had invested

in tools to address that concern.

89

A separate analysis found that 57 per

cent of organisations acknowledged

90

that AI was advancing more

quickly than they could secure it.

91

The phrase "governance gap" has

become a staple of policy discourse.

92

It undersells the scale of the problem.

93

This is not a gap.

94

It is a chasm.

95

The Partnership on AI identified

96

responsible adoption of agentic AI

systems, improved documentation and

97

transparency standards, governance

convergence across jurisdictions,

98

protections for authentic human

voice in an era of synthetic content.

99

The priorities are sensible.

100

They are also an implicit admission

that none of these foundations are yet

101

in place, despite years of discussion.

102

Meanwhile, agentic AI systems — which

take autonomous actions in the real

103

world rather than simply generating

text — introduce what the Partnership

104

describes as non-reversibility of

actions, open-ended decision-making

105

pathways, and privacy vulnerabilities

from expanded data access.

106

These are not theoretical risks.

107

They are features of systems already

deployed in customer service, software

108

development, and financial trading.

109

The governance frameworks meant

to constrain them are, in many

110

cases, still being drafted.

111

The European Union's AI Act

represents the most ambitious

112

attempt to date to translate ethical

principles into enforceable law.

113

It entered into force in August

114

timeline extending through 2027.

115

Prohibitions on the most dangerous AI

116

Full enforcement of requirements for

high-risk systems — with fines reaching

117

up to 35 million euros or seven per

cent of global annual turnover — does

118

not arrive until August 2026.

119

This is, by any measure, a

significant regulatory achievement.

120

But the Act was first

121

When the European Commission drafted

that proposal, ChatGPT did not exist.

122

Nor did the current generation of

autonomous agents, multimodal models,

123

or AI-powered code generation tools.

124

The regulation is, by design,

chasing a target that moved while

125

lawmakers were still aiming.

126

The United States presents a

different set of challenges entirely.

127

Rather than pursuing comprehensive

legislation, it has relied on a

128

decentralised approach combining

agency-specific enforcement, voluntary

129

frameworks, and sector-level regulation.

130

Then, in December 2025, President Trump

signed an executive order seeking what the

131

administration described as a minimally

burdensome national policy framework.

132

The order directed the Attorney

General to establish an AI Litigation

133

Task Force to challenge state AI

laws deemed inconsistent with federal

134

policy, and instructed the Secretary

of Commerce to identify state

135

legislation considered "onerous."

136

It even tied federal broadband

infrastructure funding to compliance

137

with those determinations.

138

The order was, in effect, an

attempt to pre-empt a patchwork of

139

state-level regulations that had

been emerging with genuine ambition.

140

Colorado's legislation, effective February

141

of high-risk AI systems to use reasonable

care to protect consumers from algorithmic

142

discrimination, implement risk management

policies, and conduct impact assessments.

143

New York City had already established

bias audit requirements for

144

automated employment decision tools.

145

More than a hundred state AI

laws were enacted across the

146

United States in 2025 alone.

147

Governors in California, Colorado,

and New York indicated they would

148

enforce their statutes regardless.

149

Legal scholars noted the constitutional

questions were substantial.

150

The result is a governance landscape that

is not merely fragmented but actively

151

contested, with federal and state

authorities pulling in opposing directions

152

whilst companies navigate overlapping

and sometimes contradictory obligations.

153

When enforcement mechanisms fail to

materialise in time, the costs do

154

not distribute themselves evenly.

155

They concentrate, with brutal

predictability, on those with

156

the least power to resist.

157

In employment, five individuals

over the age of forty applied

158

for hundreds of positions through

Workday's automated hiring platform

159

and received almost no interviews.

160

They alleged that Workday's

AI recommendation system

161

discriminated on the basis of age.

162

In 2024, a court allowed the disparate

impact claim to proceed, holding that

163

Workday bore liability as an agent

of the employers using its product.

164

In housing, plaintiffs demonstrated

that the SafeRent tenant screening

165

algorithm produced discriminatory

outcomes for Black and Hispanic

166

applicants, and the company settled for

167

In biometric surveillance, Clearview

AI scraped billions of photographs

168

from social media without consent,

sold facial recognition services to law

169

enforcement worldwide, was fined 30.5

170

million euros by the Dutch data

protection authority, and then settled a

171

US class action for approximately 51.75

172

million dollars — structured,

extraordinarily, as a 23 per

173

cent equity stake in the company

itself, because Clearview had

174

insufficient assets to pay in cash.

175

A bipartisan group of state attorneys

general filed formal objections

176

to the settlement's adequacy.

177

These cases share a common structure.

178

Harm occurs.

179

Years pass.

180

Legal proceedings unfold.

181

Settlements are reached or fines imposed.

182

But the systems that caused the harm

often continue operating throughout

183

the entire adjudication process, and

the individuals affected rarely receive

184

compensation proportional to their injury.

185

The enforcement mechanisms

exist, technically.

186

They simply do not work fast

enough to prevent the damage

187

they are meant to address.

188

A study from the University of Washington

provided evidence of the scale of

189

algorithmic bias in employment contexts.

190

Researchers presented three AI models

with job applications identical in every

191

respect except the name of the applicant.

192

The models preferred resumes with

white-associated names in 85 per cent

193

of cases and those with Black-associated

names only 9 per cent of the time.

194

A separate study published in June 2025,

led by researchers at Cedars-Sinai,

195

found that leading large language models

generated less effective treatment

196

recommendations when a patient's race

was identified as African American.

197

These are not edge cases.

198

They are documented patterns

of discriminatory behaviour

199

embedded in systems that millions

of people interact with daily.

200

And they persist not because the ethical

principles are inadequate, but because

201

the mechanisms for enforcing those

principles remain woefully underdeveloped.

202

Algorithmic auditing is often proposed

as the solution: independent third

203

parties evaluating AI systems for

bias and compliance, much as financial

204

auditors examine corporate accounts.

205

New York City requires annual bias audits

for automated employment decision tools.

206

Colorado mandates impact

assessments for high-risk systems.

207

The EU AI Act requires conformity

assessments for high-risk applications.

208

The AI Now Institute has mounted a

detailed critique of this approach,

209

arguing that technical evaluations

narrowly position bias as a flaw that

210

can be fixed and eliminated, when in fact

algorithmic harms are often structural,

211

reflecting the social contexts in which

systems are designed and deployed.

212

Audits, the institute contends,

risk entrenching power within the

213

technology industry whilst taking focus

away from more structural responses.

214

The critique has substance.

215

There are no universally

accepted standards for what

216

constitutes a passing score.

217

Audit costs range from five

thousand to fifty thousand pounds

218

depending on system complexity.

219

Audits evaluate systems at a single

point in time, but AI models drift

220

as they encounter new data, meaning a

system that passes today may produce

221

discriminatory outcomes next month.

222

And audits place the primary

burden of accountability on

223

those with the fewest resources.

224

The information asymmetry is

profound and, under current

225

frameworks, largely unaddressed.

226

The geopolitical dimension

complicates everything further.

227

At the AI Action Summit

228

58 nations signed a joint

declaration on inclusive and

229

sustainable artificial intelligence.

230

The United States and the

United Kingdom refused to sign.

231

Anthropic's chief executive described

the summit as a missed opportunity

232

for addressing AI safety, reflecting a

broader frustration that international

233

forums produce declarations

rather than binding commitments.

234

Meanwhile, researchers have documented

how AI development mirrors historical

235

patterns of colonial resource extraction.

236

Control over data infrastructures,

computational resources, and

237

algorithmic systems remains

concentrated in a small number of

238

wealthy nations and corporations.

239

Environmental costs fall

disproportionately on regions where data

240

centres proliferate because electricity

and land are cheap, exporting the

241

benefits of artificial intelligence

whilst localising its burdens.

242

When 98 per cent of AI research originates

from wealthy institutions, the resulting

243

systems embed assumptions that may

be irrelevant or damaging elsewhere.

244

The world is not building

a shared technology.

245

It is building one that reflects

the interests of those who built it.

246

What would a more effective

system actually require?

247

The evidence points to several structural

necessities that go beyond the familiar

248

call for more principles or better audits.

249

Accountability must be anticipatory

rather than reactive — the current

250

model waits for harm to occur, then

attempts to assign responsibility through

251

litigation or regulatory action long

after damage has accumulated across

252

thousands of individual decisions.

253

Enforcement must be resourced

proportionally to the scale of deployment.

254

The finding that only 31 per cent

of organisations have comprehensive

255

AI policies is not simply a failure

of corporate governance; it reflects

256

a reality in which the institutions

responsible for oversight lack the

257

funding, technical expertise, and legal

authority to match the pace of industry.

258

Transparency must extend beyond model

documentation to encompass the full chain

259

of development and deployment, enabling

affected communities — not just regulators

260

and auditors — to understand how decisions

are made and what recourse is available.

261

The costs of non-compliance

must be sufficiently high to

262

alter corporate behaviour.

263

And governance frameworks must be

designed for iteration, not permanence,

264

because a five-year legislative cycle

is simply incompatible with a technology

265

that transforms every six months.

266

None of these requirements are novel.

267

Researchers, civil society

organisations, and some regulators have

268

been advocating for them for years.

269

The obstacle is not a lack of ideas.

270

It is a lack of political will,

complicated by the enormous economic

271

interests that benefit from the

current arrangement — in which

272

deployment runs ahead of governance,

and the costs of failure are borne by

273

those least equipped to absorb them.

274

The financial scale of what

has been allowed to occur

275

is staggering in aggregate.

276

Individual settlements and fines

may appear substantial in isolation.

277

Set against the revenues of the companies

deploying these systems, and against

278

the cumulative harm inflicted across

millions of affected individuals, they

279

represent a cost of doing business

rather than a meaningful deterrent.

280

The economics of non-compliance

remain, for the moment, firmly

281

in favour of deploying first

and accounting for it later.

282

The field of AI ethics has accomplished

something genuinely remarkable in building

283

global consensus around core values.

284

That achievement should not be dismissed.

285

But consensus without enforcement

is aspiration without consequence.

286

And aspiration without consequence

is just another way of saying

287

that nobody is responsible.