Watermarking large language models in Europe: interpreting the AI act in light of technology

{"Abstract":[0],"To":[1],"foster":[2],"trustworthy":[3],"Artificial":[4],"Intelligence":[5],"(AI)":[6],"within":[7,226],"the":[8,11,20,45,81,87,117,120,145,162,168,199,227],"European":[9,84,201],"Union,":[10],"AI":[12,147],"Act":[13],"requires":[14],"providers":[15],"to":[16,61,116,184],"mark":[17],"and":[18,29,42,48,72,94,134,159,164,203],"detect":[19],"outputs":[21],"of":[22,51,83,89,112,119,161,165,167,230],"their":[23],"general-purpose":[24],"models.":[25],"The":[26],"Article":[27],"50":[28],"Recital":[30],"133":[31],"call":[32],"for":[33,53,196],"marking":[34],"methods":[35,114,195],"that":[36,205],"are":[37,126],"\\"sufficiently":[38],"reliable,":[39],"interoperable,":[40],"effective":[41],"robust\\".":[43],"Yet,":[44],"rapidly":[46],"evolving":[47],"heterogeneous":[49],"landscape":[50],"watermarks":[52],"Large":[54],"Language":[55],"Models":[56],"(LLMs)":[57],"makes":[58],"it":[59],"difficult":[60],"determine":[62],"how":[63],"these":[64],"four":[65,211],"standards":[66],"can":[67],"be":[68],"translated":[69],"into":[70,222],"concrete":[71],"measurable":[73],"evaluations.":[74],"Our":[75],"paper":[76],"addresses":[77],"this":[78],"challenge,":[79],"anchoring":[80],"normativity":[82],"requirements":[85,149],"in":[86,175],"multiplicity":[88],"watermarking":[90,113,177,194,223],"techniques.":[91],"Introducing":[92],"clear":[93],"distinct":[95],"concepts":[96],"on":[97,157],"LLM":[98,121,176],"watermarking,":[99],"our":[100],"contribution":[101],"is":[102],"threefold:":[103],"(i)":[104],"Watermarking":[105,141,189],"Categorisation:":[106],"We":[107,143,191],"propose":[108,180],"an":[109],"accessible":[110],"taxonomy":[111],"according":[115],"stage":[118],"lifecycle":[122],"at":[123],"which":[124],"they":[125],"applied":[127],"—":[128],"before,":[129],"during,":[130],"or":[131,138],"after":[132],"training,":[133],"during":[135],"next-token":[136],"distribution":[137],"sampling.":[139],"(ii)":[140],"Evaluation:":[142],"interpret":[144],"EU":[146],"Act’s":[148],"by":[150,214],"mapping":[151],"each":[152],"criterion":[153],"with":[154],"state-of-the-art":[155],"evaluations":[156],"robustness":[158],"detectability":[160],"watermark,":[163],"quality":[166],"LLM.":[169],"Since":[170],"interoperability":[171],"remains":[172],"largely":[173],"untheorised":[174],"research,":[178],"we":[179,218],"three":[181],"normative":[182],"dimensions":[183],"frame":[185],"its":[186],"assessment.":[187],"(iii)":[188],"Comparison:":[190],"compare":[192],"current":[193],"LLMs":[197],"against":[198],"operationalised":[200],"criteria":[202],"show":[204],"no":[206],"approach":[207],"yet":[208],"satisfies":[209],"all":[210],"standards.":[212],"Encouraged":[213],"emerging":[215],"empirical":[216],"tests,":[217],"recommend":[219],"further":[220],"research":[221],"directly":[224],"embedded":[225],"low-level":[228],"architecture":[229],"LLMs.":[231]}