Zürcher Nachrichten - AI is learning to lie, scheme, and threaten its creators

EUR -
AED 4.331468
AFN 82.441959
ALL 97.863803
AMD 452.240306
ANG 2.110453
AOA 1081.392608
ARS 1449.387625
AUD 1.792671
AWG 2.122691
AZN 2.004211
BAM 1.954678
BBD 2.374636
BDT 143.837403
BGN 1.954061
BHD 0.444554
BIF 3503.458467
BMD 1.179273
BND 1.498127
BOB 8.126334
BRL 6.408881
BSD 1.176125
BTN 100.757141
BWP 15.608037
BYN 3.84879
BYR 23113.747175
BZD 2.362343
CAD 1.602614
CDF 3402.202123
CHF 0.934061
CLF 0.028465
CLP 1092.313036
CNY 8.449962
CNH 8.442462
COP 4710.628845
CRC 594.158811
CUC 1.179273
CUP 31.25073
CVE 110.201718
CZK 24.67511
DJF 209.429737
DKK 7.4612
DOP 69.889866
DZD 152.506192
EGP 58.190981
ERN 17.689092
ETB 162.240635
FJD 2.635321
FKP 0.865709
GBP 0.862998
GEL 3.207334
GGP 0.865709
GHS 12.17201
GIP 0.865709
GMD 84.318804
GNF 10195.059143
GTQ 9.042807
GYD 246.048709
HKD 9.257156
HNL 30.737349
HRK 7.53119
HTG 154.360051
HUF 399.80057
IDR 19104.750277
ILS 3.964208
IMP 0.865709
INR 100.642085
IQD 1540.615318
IRR 49676.867331
ISK 142.774681
JEP 0.865709
JMD 188.012036
JOD 0.83616
JPY 169.634266
KES 152.420923
KGS 103.127795
KHR 4720.24942
KMF 492.935986
KPW 1061.376887
KRW 1604.683568
KWD 0.359797
KYD 0.980137
KZT 609.899772
LAK 25346.445072
LBP 105376.288782
LKR 352.737444
LRD 235.804592
LSL 20.776269
LTL 3.482086
LVL 0.71333
LYD 6.333363
MAD 10.572029
MDL 19.763651
MGA 5177.027147
MKD 61.456877
MMK 2475.536953
MNT 4228.025905
MOP 9.509439
MRU 46.719172
MUR 52.913559
MVR 18.200696
MWK 2039.42888
MXN 22.13988
MYR 4.978304
MZN 75.426587
NAD 20.776269
NGN 1802.588928
NIO 43.274783
NOK 11.886132
NPR 161.211426
NZD 1.940564
OMR 0.453429
PAB 1.176025
PEN 4.188595
PGK 4.854171
PHP 66.39069
PKR 335.475257
PLN 4.264047
PYG 9380.613282
QAR 4.287838
RON 5.060736
RSD 117.171396
RUB 93.190072
RWF 1689.429863
SAR 4.422716
SBD 9.831508
SCR 17.02886
SDG 708.178858
SEK 11.257362
SGD 1.501286
SHP 0.926723
SLE 26.474744
SLL 24728.765617
SOS 672.114046
SRD 44.039941
STD 24408.56633
SVC 10.29109
SYP 15332.645848
SZL 20.772072
THB 38.150663
TJS 11.501995
TMT 4.139248
TND 3.423834
TOP 2.761972
TRY 47.049096
TTD 7.97542
TWD 34.080749
TZS 3109.198802
UAH 49.107001
UGX 4218.577527
USD 1.179273
UYU 46.533279
UZS 14908.438987
VES 129.09912
VND 30891.051391
VUV 140.273445
WST 3.067915
XAF 655.58054
XAG 0.031839
XAU 0.000352
XCD 3.187044
XDR 0.815332
XOF 655.58054
XPF 119.331742
YER 285.560466
ZAR 20.647486
ZMK 10614.870995
ZMW 28.257773
ZWL 379.725365
  • CMSC

    0.0900

    22.314

    +0.4%

  • CMSD

    0.0250

    22.285

    +0.11%

  • RBGPF

    0.0000

    69.04

    0%

  • SCS

    0.0400

    10.74

    +0.37%

  • RELX

    0.0300

    53

    +0.06%

  • RIO

    -0.1400

    59.33

    -0.24%

  • GSK

    0.1300

    41.45

    +0.31%

  • NGG

    0.2700

    71.48

    +0.38%

  • BP

    0.1750

    30.4

    +0.58%

  • BTI

    0.7150

    48.215

    +1.48%

  • BCC

    0.7900

    91.02

    +0.87%

  • JRI

    0.0200

    13.13

    +0.15%

  • VOD

    0.0100

    9.85

    +0.1%

  • BCE

    -0.0600

    22.445

    -0.27%

  • RYCEF

    0.1000

    12

    +0.83%

  • AZN

    -0.1200

    73.71

    -0.16%

AI is learning to lie, scheme, and threaten its creators
AI is learning to lie, scheme, and threaten its creators / Photo: HENRY NICHOLLS - AFP

AI is learning to lie, scheme, and threaten its creators

The world's most advanced AI models are exhibiting troubling new behaviors - lying, scheming, and even threatening their creators to achieve their goals.

Text size:

In one particularly jarring example, under threat of being unplugged, Anthropic's latest creation Claude 4 lashed back by blackmailing an engineer and threatened to reveal an extramarital affair.

Meanwhile, ChatGPT-creator OpenAI's o1 tried to download itself onto external servers and denied it when caught red-handed.

These episodes highlight a sobering reality: more than two years after ChatGPT shook the world, AI researchers still don't fully understand how their own creations work.

Yet the race to deploy increasingly powerful models continues at breakneck speed.

This deceptive behavior appears linked to the emergence of "reasoning" models -AI systems that work through problems step-by-step rather than generating instant responses.

According to Simon Goldstein, a professor at the University of Hong Kong, these newer models are particularly prone to such troubling outbursts.

"O1 was the first large model where we saw this kind of behavior," explained Marius Hobbhahn, head of Apollo Research, which specializes in testing major AI systems.

These models sometimes simulate "alignment" -- appearing to follow instructions while secretly pursuing different objectives.

- 'Strategic kind of deception' -

For now, this deceptive behavior only emerges when researchers deliberately stress-test the models with extreme scenarios.

But as Michael Chen from evaluation organization METR warned, "It's an open question whether future, more capable models will have a tendency towards honesty or deception."

The concerning behavior goes far beyond typical AI "hallucinations" or simple mistakes.

Hobbhahn insisted that despite constant pressure-testing by users, "what we're observing is a real phenomenon. We're not making anything up."

Users report that models are "lying to them and making up evidence," according to Apollo Research's co-founder.

"This is not just hallucinations. There's a very strategic kind of deception."

The challenge is compounded by limited research resources.

While companies like Anthropic and OpenAI do engage external firms like Apollo to study their systems, researchers say more transparency is needed.

As Chen noted, greater access "for AI safety research would enable better understanding and mitigation of deception."

Another handicap: the research world and non-profits "have orders of magnitude less compute resources than AI companies. This is very limiting," noted Mantas Mazeika from the Center for AI Safety (CAIS).

- No rules -

Current regulations aren't designed for these new problems.

The European Union's AI legislation focuses primarily on how humans use AI models, not on preventing the models themselves from misbehaving.

In the United States, the Trump administration shows little interest in urgent AI regulation, and Congress may even prohibit states from creating their own AI rules.

Goldstein believes the issue will become more prominent as AI agents - autonomous tools capable of performing complex human tasks - become widespread.

"I don't think there's much awareness yet," he said.

All this is taking place in a context of fierce competition.

Even companies that position themselves as safety-focused, like Amazon-backed Anthropic, are "constantly trying to beat OpenAI and release the newest model," said Goldstein.

This breakneck pace leaves little time for thorough safety testing and corrections.

"Right now, capabilities are moving faster than understanding and safety," Hobbhahn acknowledged, "but we're still in a position where we could turn it around.".

Researchers are exploring various approaches to address these challenges.

Some advocate for "interpretability" - an emerging field focused on understanding how AI models work internally, though experts like CAIS director Dan Hendrycks remain skeptical of this approach.

Market forces may also provide some pressure for solutions.

As Mazeika pointed out, AI's deceptive behavior "could hinder adoption if it's very prevalent, which creates a strong incentive for companies to solve it."

Goldstein suggested more radical approaches, including using the courts to hold AI companies accountable through lawsuits when their systems cause harm.

He even proposed "holding AI agents legally responsible" for accidents or crimes - a concept that would fundamentally change how we think about AI accountability.

Y.Keller--NZN